From b138a41e26c47e32a0212973c639a78743752433 Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-37-72.us-east-2.compute.internal>
Date: Thu, 10 Mar 2022 14:56:23 +0000
Subject: [PATCH 01/22] add path to db

---
 egs2/TEMPLATE/asr1/db.sh | 1 +
 1 file changed, 1 insertion(+)

diff --git a/egs2/TEMPLATE/asr1/db.sh b/egs2/TEMPLATE/asr1/db.sh
index 3785aef57a8..b0d2b9ccb8f 100755
--- a/egs2/TEMPLATE/asr1/db.sh
+++ b/egs2/TEMPLATE/asr1/db.sh
@@ -126,6 +126,7 @@ PRIMEWORDS_CHINESE=downloads
 SEAME=
 BENGALI=downloads
 IWSLT14=
+ST_CMDS=downloads
 
 # For only CMU TIR environment
 if [[ "$(hostname)" == tir* ]]; then

From 11bbacbde827cf51cbe916f3126b8ea796d8fedb Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-37-72.us-east-2.compute.internal>
Date: Thu, 10 Mar 2022 14:57:35 +0000
Subject: [PATCH 02/22] initial commit

---
 egs2/zh_openslr38/asr1/README.md              | 179 ++++++++++++++++++
 egs2/zh_openslr38/asr1/asr.sh                 |   1 +
 egs2/zh_openslr38/asr1/cmd.sh                 | 110 +++++++++++
 .../asr1/conf/decode_asr_rnn.yaml             |   6 +
 .../asr1/conf/decode_asr_simu_streaming.yaml  |   7 +
 .../asr1/conf/decode_asr_streaming.yaml       |  11 ++
 .../asr1/conf/decode_asr_transformer.yaml     |   6 +
 .../conf/decode_asr_transformer_ngram.yaml    |   7 +
 egs2/zh_openslr38/asr1/conf/fbank.conf        |   2 +
 egs2/zh_openslr38/asr1/conf/pbs.conf          |  11 ++
 egs2/zh_openslr38/asr1/conf/pitch.conf        |   1 +
 egs2/zh_openslr38/asr1/conf/queue.conf        |  12 ++
 egs2/zh_openslr38/asr1/conf/slurm.conf        |  14 ++
 .../asr1/conf/train_asr_conformer.yaml        |   1 +
 .../zh_openslr38/asr1/conf/train_asr_rnn.yaml |  55 ++++++
 .../conf/train_asr_streaming_conformer.yaml   |  95 ++++++++++
 .../conf/train_asr_streaming_transformer.yaml |  76 ++++++++
 .../asr1/conf/train_asr_transformer.yaml      |   1 +
 egs2/zh_openslr38/asr1/conf/train_lm.yaml     |  19 ++
 .../asr1/conf/train_lm_transformer.yaml       |   1 +
 .../asr1/conf/tuning/train_asr_conformer.yaml |  76 ++++++++
 .../conf/tuning/train_asr_transformer.yaml    |  66 +++++++
 .../conf/tuning/train_asr_transformer2.yaml   |  54 ++++++
 .../conf/tuning/train_asr_transformer3.yaml   |  70 +++++++
 .../tuning/train_asr_transformer_lr0.002.yaml |  66 +++++++
 .../asr1/conf/tuning/train_lm_rnn.yaml        |  19 ++
 .../asr1/conf/tuning/train_lm_rnn2.yaml       |  19 ++
 .../conf/tuning/train_lm_transformer.yaml     |  29 +++
 egs2/zh_openslr38/asr1/db.sh                  |   1 +
 egs2/zh_openslr38/asr1/local/data.sh          |  85 +++++++++
 egs2/zh_openslr38/asr1/local/data_download.sh |  50 +++++
 egs2/zh_openslr38/asr1/local/data_split.py    |  97 ++++++++++
 .../asr1/local/download_and_untar.sh          |   1 +
 egs2/zh_openslr38/asr1/local/path.sh          |   1 +
 egs2/zh_openslr38/asr1/path.sh                |   1 +
 egs2/zh_openslr38/asr1/pyscripts              |   1 +
 egs2/zh_openslr38/asr1/run.sh                 |  40 ++++
 egs2/zh_openslr38/asr1/run_streaming.sh       |  41 ++++
 egs2/zh_openslr38/asr1/scripts                |   1 +
 egs2/zh_openslr38/asr1/steps                  |   1 +
 egs2/zh_openslr38/asr1/utils                  |   1 +
 41 files changed, 1335 insertions(+)
 create mode 100644 egs2/zh_openslr38/asr1/README.md
 create mode 120000 egs2/zh_openslr38/asr1/asr.sh
 create mode 100644 egs2/zh_openslr38/asr1/cmd.sh
 create mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_rnn.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_simu_streaming.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_streaming.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_transformer.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_transformer_ngram.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/fbank.conf
 create mode 100644 egs2/zh_openslr38/asr1/conf/pbs.conf
 create mode 100644 egs2/zh_openslr38/asr1/conf/pitch.conf
 create mode 100644 egs2/zh_openslr38/asr1/conf/queue.conf
 create mode 100644 egs2/zh_openslr38/asr1/conf/slurm.conf
 create mode 120000 egs2/zh_openslr38/asr1/conf/train_asr_conformer.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/train_asr_rnn.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/train_asr_streaming_conformer.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/train_asr_streaming_transformer.yaml
 create mode 120000 egs2/zh_openslr38/asr1/conf/train_asr_transformer.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/train_lm.yaml
 create mode 120000 egs2/zh_openslr38/asr1/conf/train_lm_transformer.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_conformer.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer2.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer3.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer_lr0.002.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn2.yaml
 create mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_lm_transformer.yaml
 create mode 120000 egs2/zh_openslr38/asr1/db.sh
 create mode 100755 egs2/zh_openslr38/asr1/local/data.sh
 create mode 100755 egs2/zh_openslr38/asr1/local/data_download.sh
 create mode 100644 egs2/zh_openslr38/asr1/local/data_split.py
 create mode 120000 egs2/zh_openslr38/asr1/local/download_and_untar.sh
 create mode 100755 egs2/zh_openslr38/asr1/local/path.sh
 create mode 120000 egs2/zh_openslr38/asr1/path.sh
 create mode 120000 egs2/zh_openslr38/asr1/pyscripts
 create mode 100755 egs2/zh_openslr38/asr1/run.sh
 create mode 100755 egs2/zh_openslr38/asr1/run_streaming.sh
 create mode 120000 egs2/zh_openslr38/asr1/scripts
 create mode 120000 egs2/zh_openslr38/asr1/steps
 create mode 120000 egs2/zh_openslr38/asr1/utils

diff --git a/egs2/zh_openslr38/asr1/README.md b/egs2/zh_openslr38/asr1/README.md
new file mode 100644
index 00000000000..985604c1a43
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/README.md
@@ -0,0 +1,179 @@
+# Streaming Conformer + specaug + speed perturbation: feats=raw, n_fft=512, hop_length=128
+## Environments
+- date: `Mon Aug 23 16:31:48 CST 2021`
+- python version: `3.7.9 (default, Aug 31 2020, 12:42:55)  [GCC 7.3.0]`
+- espnet version: `espnet 0.9.9`
+- pytorch version: `pytorch 1.5.0`
+- Git hash: `b94d07028099a80c9c690341981ae7d550b5ca24`
+  - Commit date: `Mon Aug 23 00:47:47 2021 +0800`
+
+## With Transformer LM
+- Model link: (wait for upload)
+- ASR config: [./conf/train_asr_streaming_cpnformer.yaml](./conf/train_asr_streaming_conformer.yaml)
+- LM config: [./conf/tuning/train_lm_transformer.yaml](./conf/tuning/train_lm_transformer.yaml)
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_streaming_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.0|5.8|0.3|0.3|6.3|42.2|
+|decode_asr_streaming_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|92.9|6.7|0.5|0.7|7.8|46.2|
+# Streaming Transformer + speed perturbation: feats=raw, n_fft=512, hop_length=128
+## Environments
+- date: `Tue Aug 17 01:20:32 CST 2021`
+- python version: `3.7.9 (default, Aug 31 2020, 12:42:55)  [GCC 7.3.0]`
+- espnet version: `espnet 0.9.9`
+- pytorch version: `pytorch 1.5.0`
+- Git hash: `6f5f848e0a9bfca1b73393779233bde34add3df1`
+  - Commit date: `Mon Aug 16 21:50:08 2021 +0800`
+
+## With Transformer LM
+- Model link: (wait for upload)
+- ASR config: [./conf/train_asr_streaming_transformer.yaml](./conf/train_asr_streaming_transformer.yaml)
+- LM config: [./conf/tuning/train_lm_transformer.yaml](./conf/tuning/train_lm_transformer.yaml)
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_streaming_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|93.6|6.2|0.1|0.5|6.8|46.8|
+|decode_asr_streaming_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.0|6.7|0.2|0.8|7.8|50.7|
+
+# Conformer + specaug + speed perturbation: feats=raw, n_fft=512, hop_length=128
+## Environments
+- date: `Fri Oct 16 11:10:17 JST 2020`
+- python version: `3.7.3 (default, Mar 27 2019, 22:11:17)  [GCC 7.3.0]`
+- espnet version: `espnet 0.9.0`
+- pytorch version: `pytorch 1.6.0`
+- Git hash: `20b0c89369d9dd3e05780b65fdd00a9b4f4891e5`
+  - Commit date: `Mon Oct 12 09:28:20 2020 -0400`
+
+## With Transformer LM
+- Model link: https://zenodo.org/record/4105763#.X40xe2j7QUE
+- ASR config: [./conf/tuning/train_asr_conformer.yaml](./conf/tuning/train_asr_conformer.yaml)
+- LM config: [./conf/tuning/train_lm_transformer.yaml](./conf/tuning/train_lm_transformer.yaml)
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_rnn_lm_lm_train_lm_transformer_char_batch_bins2000000_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|95.7|4.2|0.1|0.1|4.4|33.7|
+|decode_asr_rnn_lm_lm_train_lm_transformer_char_batch_bins2000000_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|95.4|4.5|0.1|0.1|4.7|35.0|
+
+## With RNN LM
+- ASR config: [./conf/tuning/train_asr_conformer.yaml](./conf/tuning/train_asr_conformer.yaml)
+- LM config: [./conf/tuning/train_lm_rnn2.yaml](./conf/tuning/train_lm_rnn2.yaml)
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|95.5|4.4|0.1|0.1|4.6|35.2|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|95.2|4.7|0.1|0.1|4.9|36.5|
+
+## Without LM
+- ASR config: [./conf/tuning/train_asr_conformer.yaml](./conf/tuning/train_asr_conformer.yaml)
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_rnn_asr_model_valid.acc.ave/dev|14326|205341|95.6|4.3|0.1|0.1|4.5|35.0|
+|decode_asr_rnn_asr_model_valid.acc.ave/test|7176|104765|95.2|4.7|0.1|0.1|4.9|36.7|
+
+# Transformer + speed perturbation: feats=raw with same LM with the privious setting
+
+I compared between `n_fft=512, hop_length=128`, `n_fft=400, hop_length=160`,  and `n_fft=512, hop_length=256`
+with searching the best `batch_bins` to get the suitable configuration for each hop_length.
+
+## Environments
+- date: `Fri Oct 16 11:10:17 JST 2020`
+- python version: `3.7.3 (default, Mar 27 2019, 22:11:17)  [GCC 7.3.0]`
+- espnet version: `espnet 0.9.0`
+- pytorch version: `pytorch 1.6.0`
+- Git hash: `20b0c89369d9dd3e05780b65fdd00a9b4f4891e5`
+  - Commit date: `Mon Oct 12 09:28:20 2020 -0400`
+
+## n_fft=512, hop_length=128
+asr_train_asr_transformer2_raw_char_batch_typenumel_batch_bins8500000_optim_conflr0.0005_scheduler_confwarmup_steps30000_sp
+
+- ASR config: [./conf/tuning/train_asr_transformer3.yaml](./conf/tuning/train_asr_transformer3.yaml)
+- LM config: [./conf/tuning/train_lm_rnn2.yaml](./conf/tuning/train_lm_rnn2.yaml)
+
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.2|5.7|0.1|0.1|5.9|42.6|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.7|6.1|0.2|0.1|6.4|45.0|
+
+
+## n_fft=400, hop_length=160
+asr_train_asr_transformer2_raw_char_frontend_confn_fft400_frontend_confhop_length160_batch_typenumel_batch_bins6500000_optim_conflr0.0005_scheduler_confwarmup_steps30000_sp
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.1|5.7|0.1|0.1|6.0|43.0|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.5|6.3|0.2|0.1|6.6|45.4|
+
+## n_fft=512, hop_length=256
+asr_train_asr_transformer2_raw_char_frontend_confn_fft512_frontend_confhop_length256_batch_typenumel_batch_bins6000000_sp
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.0|5.9|0.1|0.1|6.1|43.5|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.3|6.5|0.2|0.1|6.8|45.8|
+
+
+# Transformer + speed perturbation: feats=fbank_pitch, RNN-LM
+Compatible setting with espnet1 to reproduce the previou result
+
+## Environments
+- date: `Fri Oct 16 11:10:17 JST 2020`
+- python version: `3.7.3 (default, Mar 27 2019, 22:11:17)  [GCC 7.3.0]`
+- espnet version: `espnet 0.9.0`
+- pytorch version: `pytorch 1.6.0`
+- Git hash: `20b0c89369d9dd3e05780b65fdd00a9b4f4891e5`
+  - Commit date: `Mon Oct 12 09:28:20 2020 -0400`
+
+- ASR config: [./conf/tuning/train_asr_transformer2.yaml](./conf/tuning/train_asr_transformer2.yaml)
+- LM config: [./conf/tuning/train_lm_rnn2.yaml](./conf/tuning/train_lm_rnn2.yaml)
+
+## asr_train_asr_transformer2_fbank_pitch_char_sp
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.0|5.9|0.1|0.1|6.1|43.4|
+|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.4|6.4|0.2|0.1|6.7|45.9|
+
+# The first result
+## Environments
+- date: `Sun Feb  2 02:03:55 CST 2020`
+- python version: `3.7.3 (default, Mar 27 2019, 22:11:17)  [GCC 7.3.0]`
+- espnet version: `espnet 0.6.0`
+- pytorch version: `pytorch 1.1.0`
+- Git hash: `e0fd073a70bcded6a0e6a3587630410a994ccdb8` (+ fixing https://github.com/espnet/espnet/pull/1533)
+  - Commit date: `Sat Jan 11 06:09:24 2020 +0900`
+
+## asr_train_asr_rnn_new_fbank_pitch_char
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_devdecode_asr_rnn_lm_valid.loss.best_asr_model_valid.acc.best|14326|205341|92.6|7.2|0.2|0.1|7.5|49.6|
+|decode_testdecode_asr_rnn_lm_valid.loss.best_asr_model_valid.acc.best|7176|104765|91.6|8.2|0.3|0.2|8.6|53.4|
+
+## asr_train_asr_transformer_lr0.002_fbank_pitch_char
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_dev_decode_asr_rnn_lm_train_lm_char_valid.loss.best_asr_model_valid.acc.best|14326|205341|93.3|6.5|0.2|0.1|6.8|45.6|
+|decode_test_decode_asr_rnn_lm_train_lm_char_valid.loss.best_asr_model_valid.acc.best|7176|104765|92.7|7.1|0.3|0.1|7.4|47.6|
+
diff --git a/egs2/zh_openslr38/asr1/asr.sh b/egs2/zh_openslr38/asr1/asr.sh
new file mode 120000
index 00000000000..60b05122cfd
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/asr.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/asr.sh
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/cmd.sh b/egs2/zh_openslr38/asr1/cmd.sh
new file mode 100644
index 00000000000..2aae6919fef
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/cmd.sh
@@ -0,0 +1,110 @@
+# ====== About run.pl, queue.pl, slurm.pl, and ssh.pl ======
+# Usage: <cmd>.pl [options] JOB=1:<nj> <log> <command...>
+# e.g.
+#   run.pl --mem 4G JOB=1:10 echo.JOB.log echo JOB
+#
+# Options:
+#   --time <time>: Limit the maximum time to execute.
+#   --mem <mem>: Limit the maximum memory usage.
+#   -–max-jobs-run <njob>: Limit the number parallel jobs. This is ignored for non-array jobs.
+#   --num-threads <ngpu>: Specify the number of CPU core.
+#   --gpu <ngpu>: Specify the number of GPU devices.
+#   --config: Change the configuration file from default.
+#
+# "JOB=1:10" is used for "array jobs" and it can control the number of parallel jobs.
+# The left string of "=", i.e. "JOB", is replaced by <N>(Nth job) in the command and the log file name,
+# e.g. "echo JOB" is changed to "echo 3" for the 3rd job and "echo 8" for 8th job respectively.
+# Note that the number must start with a positive number, so you can't use "JOB=0:10" for example.
+#
+# run.pl, queue.pl, slurm.pl, and ssh.pl have unified interface, not depending on its backend.
+# These options are mapping to specific options for each backend and
+# it is configured by "conf/queue.conf" and "conf/slurm.conf" by default.
+# If jobs failed, your configuration might be wrong for your environment.
+#
+#
+# The official documentation for run.pl, queue.pl, slurm.pl, and ssh.pl:
+#   "Parallelization in Kaldi": http://kaldi-asr.org/doc/queue.html
+# =========================================================~
+
+
+# Select the backend used by run.sh from "local", "stdout", "sge", "slurm", or "ssh"
+cmd_backend='local'
+
+# Local machine, without any Job scheduling system
+if [ "${cmd_backend}" = local ]; then
+
+    # The other usage
+    export train_cmd="run.pl"
+    # Used for "*_train.py": "--gpu" is appended optionally by run.sh
+    export cuda_cmd="run.pl"
+    # Used for "*_recog.py"
+    export decode_cmd="run.pl"
+
+# Local machine logging to stdout and log file, without any Job scheduling system
+elif [ "${cmd_backend}" = stdout ]; then
+
+    # The other usage
+    export train_cmd="stdout.pl"
+    # Used for "*_train.py": "--gpu" is appended optionally by run.sh
+    export cuda_cmd="stdout.pl"
+    # Used for "*_recog.py"
+    export decode_cmd="stdout.pl"
+
+
+# "qsub" (Sun Grid Engine, or derivation of it)
+elif [ "${cmd_backend}" = sge ]; then
+    # The default setting is written in conf/queue.conf.
+    # You must change "-q g.q" for the "queue" for your environment.
+    # To know the "queue" names, type "qhost -q"
+    # Note that to use "--gpu *", you have to setup "complex_value" for the system scheduler.
+
+    export train_cmd="queue.pl"
+    export cuda_cmd="queue.pl"
+    export decode_cmd="queue.pl"
+
+
+# "qsub" (Torque/PBS.)
+elif [ "${cmd_backend}" = pbs ]; then
+    # The default setting is written in conf/pbs.conf.
+
+    export train_cmd="pbs.pl"
+    export cuda_cmd="pbs.pl"
+    export decode_cmd="pbs.pl"
+
+
+# "sbatch" (Slurm)
+elif [ "${cmd_backend}" = slurm ]; then
+    # The default setting is written in conf/slurm.conf.
+    # You must change "-p cpu" and "-p gpu" for the "partition" for your environment.
+    # To know the "partion" names, type "sinfo".
+    # You can use "--gpu * " by default for slurm and it is interpreted as "--gres gpu:*"
+    # The devices are allocated exclusively using "${CUDA_VISIBLE_DEVICES}".
+
+    export train_cmd="slurm.pl"
+    export cuda_cmd="slurm.pl"
+    export decode_cmd="slurm.pl"
+
+elif [ "${cmd_backend}" = ssh ]; then
+    # You have to create ".queue/machines" to specify the host to execute jobs.
+    # e.g. .queue/machines
+    #   host1
+    #   host2
+    #   host3
+    # Assuming you can login them without any password, i.e. You have to set ssh keys.
+
+    export train_cmd="ssh.pl"
+    export cuda_cmd="ssh.pl"
+    export decode_cmd="ssh.pl"
+
+# This is an example of specifying several unique options in the JHU CLSP cluster setup.
+# Users can modify/add their own command options according to their cluster environments.
+elif [ "${cmd_backend}" = jhu ]; then
+
+    export train_cmd="queue.pl --mem 2G"
+    export cuda_cmd="queue-freegpu.pl --mem 2G --gpu 1 --config conf/queue.conf"
+    export decode_cmd="queue.pl --mem 4G"
+
+else
+    echo "$0: Error: Unknown cmd_backend=${cmd_backend}" 1>&2
+    return 1
+fi
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_rnn.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_rnn.yaml
new file mode 100644
index 00000000000..88fdbc20b91
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/decode_asr_rnn.yaml
@@ -0,0 +1,6 @@
+beam_size: 20
+penalty: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+ctc_weight: 0.6
+lm_weight: 0.3
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_simu_streaming.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_simu_streaming.yaml
new file mode 100644
index 00000000000..f79545ddb2f
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/decode_asr_simu_streaming.yaml
@@ -0,0 +1,7 @@
+beam_size: 20
+penalty: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+ctc_weight: 0.6
+lm_weight: 0.3
+streaming: True
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_streaming.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_streaming.yaml
new file mode 100644
index 00000000000..fc2b3b1af5d
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/decode_asr_streaming.yaml
@@ -0,0 +1,11 @@
+beam_size: 10
+penalty: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+ctc_weight: 0.6
+lm_weight: 0.3
+sim_chunk_length: 512
+disable_repetition_detection: true
+decoder_text_length_limit: 0
+encoded_feat_length_limit: 0
+#streaming: True
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_transformer.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_transformer.yaml
new file mode 100644
index 00000000000..a147fa79dc8
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/decode_asr_transformer.yaml
@@ -0,0 +1,6 @@
+beam_size: 10
+penalty: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+ctc_weight: 0.5
+lm_weight: 0.7
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_transformer_ngram.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_transformer_ngram.yaml
new file mode 100644
index 00000000000..5e3e8e049d4
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/decode_asr_transformer_ngram.yaml
@@ -0,0 +1,7 @@
+beam_size: 10
+penalty: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+ctc_weight: 0.5
+lm_weight: 0.7
+ngram_weight: 0.9
diff --git a/egs2/zh_openslr38/asr1/conf/fbank.conf b/egs2/zh_openslr38/asr1/conf/fbank.conf
new file mode 100644
index 00000000000..82ac7bd0dbc
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/fbank.conf
@@ -0,0 +1,2 @@
+--sample-frequency=16000 
+--num-mel-bins=80
diff --git a/egs2/zh_openslr38/asr1/conf/pbs.conf b/egs2/zh_openslr38/asr1/conf/pbs.conf
new file mode 100644
index 00000000000..119509938ce
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/pbs.conf
@@ -0,0 +1,11 @@
+# Default configuration
+command qsub -V -v PATH -S /bin/bash
+option name=* -N $0
+option mem=* -l mem=$0
+option mem=0          # Do not add anything to qsub_opts
+option num_threads=* -l ncpus=$0
+option num_threads=1  # Do not add anything to qsub_opts
+option num_nodes=* -l nodes=$0:ppn=1
+default gpu=0
+option gpu=0
+option gpu=* -l ngpus=$0
diff --git a/egs2/zh_openslr38/asr1/conf/pitch.conf b/egs2/zh_openslr38/asr1/conf/pitch.conf
new file mode 100644
index 00000000000..e959a19d5b8
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/pitch.conf
@@ -0,0 +1 @@
+--sample-frequency=16000
diff --git a/egs2/zh_openslr38/asr1/conf/queue.conf b/egs2/zh_openslr38/asr1/conf/queue.conf
new file mode 100644
index 00000000000..500582fab31
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/queue.conf
@@ -0,0 +1,12 @@
+# Default configuration
+command qsub -v PATH -cwd -S /bin/bash -j y -l arch=*64*
+option name=* -N $0
+option mem=* -l mem_free=$0,ram_free=$0
+option mem=0          # Do not add anything to qsub_opts
+option num_threads=* -pe smp $0
+option num_threads=1  # Do not add anything to qsub_opts
+option max_jobs_run=* -tc $0
+option num_nodes=* -pe mpi $0  # You must set this PE as allocation_rule=1
+default gpu=0
+option gpu=0
+option gpu=* -l gpu=$0 -q g.q
diff --git a/egs2/zh_openslr38/asr1/conf/slurm.conf b/egs2/zh_openslr38/asr1/conf/slurm.conf
new file mode 100644
index 00000000000..3b229673638
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/slurm.conf
@@ -0,0 +1,14 @@
+# Default configuration
+command sbatch --export=PATH
+option name=* --job-name $0
+option time=* --time $0
+option mem=* --mem-per-cpu $0
+option mem=0
+option num_threads=* --cpus-per-task $0
+option num_threads=1 --cpus-per-task 1
+option num_nodes=* --nodes $0
+default gpu=0
+option gpu=0 -p cpu
+option gpu=* -p gpu --gres=gpu:$0 -c $0  # Recommend allocating more CPU than, or equal to the number of GPU
+# note: the --max-jobs-run option is supported as a special case
+# by slurm.pl and you don't have to handle it in the config file.
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_conformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_conformer.yaml
new file mode 120000
index 00000000000..10fea4f0971
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/train_asr_conformer.yaml
@@ -0,0 +1 @@
+tuning/train_asr_conformer.yaml
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_rnn.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_rnn.yaml
new file mode 100644
index 00000000000..eefe1d516fd
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/train_asr_rnn.yaml
@@ -0,0 +1,55 @@
+# network architecture
+# encoder related
+encoder: vgg_rnn
+encoder_conf:
+    rnn_type: lstm
+    bidirectional: True
+    use_projection: False
+    num_layers: 3
+    hidden_size: 1024
+    output_size: 1024
+    dropout: 0.0
+    in_channel: 1
+# decoder related
+decoder: rnn
+decoder_conf:
+    rnn_type: lstm
+    num_layers: 2
+    hidden_size: 1024
+    sampling_probability: 0.0   # scheduled sampling option
+    dropout: 0.0
+    att_conf:
+        adim: 1024
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.5
+    lsm_weight: 0.0     # label smoothing option
+
+# minibatch related
+batch_type: folded
+batch_size: 30
+
+# optimization related
+init: chainer
+max_epoch: 10
+optim: adadelta
+optim_conf:
+    lr: 1.0
+    rho: 0.95
+    eps: 1.0e-08
+    weight_decay: 0
+patience: 3
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 1
+scheduler: reducelronplateau
+scheduler_conf:
+    mode: max
+    factor: 0.5
+    patience: 1
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_streaming_conformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_streaming_conformer.yaml
new file mode 100644
index 00000000000..8e4fa086555
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/train_asr_streaming_conformer.yaml
@@ -0,0 +1,95 @@
+# network architecture
+# encoder related
+encoder: contextual_block_conformer    # contextual_block_conformer is the core of streaming conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+    activation_type: swish
+    macaron_style: true
+    use_cnn_module: true
+    cnn_module_kernel: 15
+    block_size: 40      # streaming configuration
+    hop_size: 16        # streaming configuration
+    look_ahead: 16      # streaming configuration
+    init_average: true  # streaming configuration
+    ctx_pos_enc: true   # streaming configuration
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+# minibatch related
+batch_type: folded
+batch_size: 128
+#valid_batch_size: 1
+# optimization related
+accum_grad: 1
+grad_clip: 5
+patience: 3
+max_epoch: 50
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+early_stopping_criterion:
+    - valid
+    - cer_ctc
+    - min
+keep_nbest_models: 10
+
+# NoamLR is deprecated. Use WarmupLR.
+# The following is equivalent setting for NoamLR:
+#
+#    optim: adam
+#    optim_conf:
+#        lr: 10.
+#    scheduler: noamlr
+#    scheduler_conf:
+#        model_size: 256
+#        warmup_steps: 25000
+#
+optim: adam
+optim_conf:
+    lr: 0.0005
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 30000
+num_att_plot: 0
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 40
+    num_time_mask: 2
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_streaming_transformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_streaming_transformer.yaml
new file mode 100644
index 00000000000..2abb8749e8d
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/train_asr_streaming_transformer.yaml
@@ -0,0 +1,76 @@
+# network architecture
+# encoder related
+encoder: contextual_block_transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+    block_size: 40
+    hop_size: 16
+    look_ahead: 16
+    init_average: true
+    ctx_pos_enc: true
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+# minibatch related
+batch_type: folded
+batch_size: 128
+#valid_batch_size: 1
+# optimization related
+accum_grad: 1
+grad_clip: 5
+patience: 3
+max_epoch: 20
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+early_stopping_criterion:
+    - valid
+    - cer_ctc
+    - min
+keep_nbest_models: 10
+
+# NoamLR is deprecated. Use WarmupLR.
+# The following is equivalent setting for NoamLR:
+#
+#    optim: adam
+#    optim_conf:
+#        lr: 10.
+#    scheduler: noamlr
+#    scheduler_conf:
+#        model_size: 256
+#        warmup_steps: 25000
+#
+optim: adam
+optim_conf:
+    lr: 0.001
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000
+num_att_plot: 0
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_transformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_transformer.yaml
new file mode 120000
index 00000000000..b9175b1a7fa
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/train_asr_transformer.yaml
@@ -0,0 +1 @@
+tuning/train_asr_transformer_lr0.002.yaml
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/conf/train_lm.yaml b/egs2/zh_openslr38/asr1/conf/train_lm.yaml
new file mode 100644
index 00000000000..26ea1e4f96e
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/train_lm.yaml
@@ -0,0 +1,19 @@
+# rnnlm related
+lm: seq_rnn
+lm_conf:
+    unit: 650
+    nlayers: 2
+
+# optimization related
+grad_clip: 5.0
+batch_type: folded
+batch_size: 64  # batch size in LM training
+max_epoch: 20   # if the data size is large, we can reduce this
+patience: 3
+optim: sgd
+
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 1
diff --git a/egs2/zh_openslr38/asr1/conf/train_lm_transformer.yaml b/egs2/zh_openslr38/asr1/conf/train_lm_transformer.yaml
new file mode 120000
index 00000000000..72276561c41
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/train_lm_transformer.yaml
@@ -0,0 +1 @@
+tuning/train_lm_transformer.yaml
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_conformer.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_conformer.yaml
new file mode 100644
index 00000000000..ab51681722f
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_conformer.yaml
@@ -0,0 +1,76 @@
+# network architecture
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+    pos_enc_layer_type: rel_pos
+    selfattention_layer_type: rel_selfattn
+    activation_type: swish
+    macaron_style: true
+    use_cnn_module: true
+    cnn_module_kernel: 15
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+# minibatch related
+batch_type: numel
+batch_bins: 4000000
+
+# optimization related
+accum_grad: 4
+grad_clip: 5
+max_epoch: 40
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 10
+
+optim: adam
+optim_conf:
+   lr: 0.0005
+scheduler: warmuplr
+scheduler_conf:
+   warmup_steps: 30000
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 40
+    num_time_mask: 2
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer.yaml
new file mode 100644
index 00000000000..1e34e982978
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer.yaml
@@ -0,0 +1,66 @@
+# network architecture
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+# minibatch related
+batch_type: folded
+batch_size: 28
+
+# optimization related
+accum_grad: 2
+grad_clip: 5
+patience: 3
+max_epoch: 20
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 10
+
+# NoamLR is deprecated. Use WarmupLR.
+# The following is equivalent setting for NoamLR:
+#
+#    optim: adam
+#    optim_conf:
+#        lr: 10.
+#    scheduler: noamlr
+#    scheduler_conf:
+#        model_size: 256
+#        warmup_steps: 25000
+#
+optim: adam
+optim_conf:
+    lr: 0.00395
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer2.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer2.yaml
new file mode 100644
index 00000000000..1ce32fff48f
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer2.yaml
@@ -0,0 +1,54 @@
+# network architecture
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+# minibatch related
+batch_type: folded
+batch_size: 32
+
+# optimization related
+accum_grad: 2
+grad_clip: 5
+max_epoch: 50
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 10
+
+optim: adam
+optim_conf:
+   lr: 0.0004
+scheduler: warmuplr
+scheduler_conf:
+   warmup_steps: 25000
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer3.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer3.yaml
new file mode 100644
index 00000000000..523c2bc8e84
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer3.yaml
@@ -0,0 +1,70 @@
+# network architecture
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+# minibatch related
+batch_type: numel
+batch_bins: 8500000
+
+# optimization related
+accum_grad: 2
+grad_clip: 5
+max_epoch: 50
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 10
+
+optim: adam
+optim_conf:
+   lr: 0.0005
+scheduler: warmuplr
+scheduler_conf:
+   warmup_steps: 30000
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 40
+    num_time_mask: 2
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer_lr0.002.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer_lr0.002.yaml
new file mode 100644
index 00000000000..7097638fcff
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer_lr0.002.yaml
@@ -0,0 +1,66 @@
+# network architecture
+# encoder related
+encoder: transformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+
+# minibatch related
+batch_type: folded
+batch_size: 32
+
+# optimization related
+accum_grad: 2
+grad_clip: 5
+patience: 3
+max_epoch: 20
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 10
+
+# NoamLR is deprecated. Use WarmupLR.
+# The following is equivalent setting for NoamLR:
+#
+#    optim: adam
+#    optim_conf:
+#        lr: 10.
+#    scheduler: noamlr
+#    scheduler_conf:
+#        model_size: 256
+#        warmup_steps: 25000
+#
+optim: adam
+optim_conf:
+    lr: 0.002
+scheduler: warmuplr     # pytorch v1.1.0+ required
+scheduler_conf:
+    warmup_steps: 25000
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn.yaml
new file mode 100644
index 00000000000..26ea1e4f96e
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn.yaml
@@ -0,0 +1,19 @@
+# rnnlm related
+lm: seq_rnn
+lm_conf:
+    unit: 650
+    nlayers: 2
+
+# optimization related
+grad_clip: 5.0
+batch_type: folded
+batch_size: 64  # batch size in LM training
+max_epoch: 20   # if the data size is large, we can reduce this
+patience: 3
+optim: sgd
+
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 1
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn2.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn2.yaml
new file mode 100644
index 00000000000..b7459681adf
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn2.yaml
@@ -0,0 +1,19 @@
+# rnnlm related
+lm: seq_rnn
+lm_conf:
+    unit: 650
+    nlayers: 2
+
+# optimization related
+grad_clip: 5.0
+batch_type: folded
+batch_size: 64  # batch size in LM training
+max_epoch: 40   # if the data size is large, we can reduce this
+patience: 3
+optim: sgd
+
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 1
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_transformer.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_transformer.yaml
new file mode 100644
index 00000000000..ace0739a939
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_transformer.yaml
@@ -0,0 +1,29 @@
+lm: transformer
+lm_conf:
+    pos_enc: null
+    embed_unit: 128
+    att_unit: 512
+    head: 8
+    unit: 2048
+    layer: 16
+    dropout_rate: 0.1
+
+# optimization related
+grad_clip: 5.0
+batch_type: numel
+batch_bins: 2000000
+accum_grad: 1
+max_epoch: 15  # 15epoch is enougth
+
+optim: adam
+optim_conf:
+   lr: 0.001
+scheduler: warmuplr
+scheduler_conf:
+   warmup_steps: 25000
+
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 10  # 10 is good.
diff --git a/egs2/zh_openslr38/asr1/db.sh b/egs2/zh_openslr38/asr1/db.sh
new file mode 120000
index 00000000000..50d86130898
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/db.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/db.sh
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/local/data.sh b/egs2/zh_openslr38/asr1/local/data.sh
new file mode 100755
index 00000000000..002756eb78c
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/local/data.sh
@@ -0,0 +1,85 @@
+#!/usr/bin/env bash
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+log() {
+    local fname=${BASH_SOURCE[1]##*/}
+    echo -e "$(date '+%Y-%m-%dT%H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+help_message=$(cat << EOF
+Usage: $0
+
+Options:
+    --remove_archive (bool): true or false
+      With remove_archive=True, the archives will be removed after being successfully downloaded and un-tarred.
+EOF
+)
+SECONDS=0
+
+# Data preparation related
+data_url=www.openslr.org/resources/38
+remove_archive=false
+download_opt=
+
+log "$0 $*"
+
+
+. ./utils/parse_options.sh
+
+. ./db.sh
+. ./path.sh
+. ./cmd.sh
+
+
+if [ $# -gt 1 ]; then
+  log "${help_message}"
+  exit 2
+fi
+
+if "$remove_archive"; then
+  download_opt="--remove-archive"
+fi
+
+if [ -z "${ST_CMDS}" ]; then
+  log "Error: \$ST_CMDS is not set in db.sh."
+  exit 2
+fi
+
+
+log "Download data to ${ST_CMDS}"
+if [ ! -d "${ST_CMDS}" ]; then
+    mkdir -p "${ST_CMDS}"
+fi
+# To absolute path
+ST_CMDS=$(cd ${ST_CMDS}; pwd)
+
+echo local/data_download.sh ${download_opt} "${ST_CMDS}" "${data_url}" ST-CMDS-20170001_1-OS.tar.gz
+local/data_download.sh ${download_opt} "${ST_CMDS}" "${data_url}" ST-CMDS-20170001_1-OS.tar.gz
+
+ST_CMDS_audio_dir=${ST_CMDS}/data_ST_CMDS/wav
+ST_CMDS_text=${ST_CMDS}/data_ST_CMDS/transcript/ST_CMDS_transcript_v0.8.txt
+
+log "Data Preparation"
+train_dir=data/train
+dev_dir=data/dev
+test_dir=data/test
+
+mkdir -p $train_dir
+mkdir -p $dev_dir
+mkdir -p $test_dir
+
+python3 local/data_split.py ${ST_CMDS}/ST-CMDS-20170001_1-OS
+
+for dir in $train_dir $dev_dir $test_dir; do
+  utils/utt2spk_to_spk2utt.pl $dir/utt2spk > $dir/spk2utt
+done
+
+# validate formats
+utils/validate_data_dir.sh --no-feats data/train
+utils/validate_data_dir.sh --no-feats data/dev
+utils/validate_data_dir.sh --no-feats data/test
+
+log "Successfully finished. [elapsed=${SECONDS}s]"
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/local/data_download.sh b/egs2/zh_openslr38/asr1/local/data_download.sh
new file mode 100755
index 00000000000..27edd73b27d
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/local/data_download.sh
@@ -0,0 +1,50 @@
+#!/usr/bin/env bash
+
+remove_archive=false
+
+if [ "$1" == --remove-archive ]; then
+  remove_archive=true
+  shift
+fi
+
+if [ $# -ne 3 ]; then
+  echo "Usage: $0 [--remove-archive] <data-base> <url-base>"
+  echo "e.g.: $0 /export/a05/xna/data www.openslr.org/resources/38"
+  echo "With --remove-archive it will remove the archive after successfully un-tarring it."
+fi
+
+data=$1
+url=$2
+
+if [ ! -d "$data" ]; then
+  echo "$0: no such directory $data"
+  exit 1;
+fi
+
+if [ -z "$url" ]; then
+  echo "$0: empty URL base."
+  exit 1;
+fi
+
+set -euo pipefail
+
+cwd=$(pwd)
+if [ ! -e "${data}/.complete" ]; then
+    mkdir -p "${data}"
+    cd "${data}" || exit 1;
+    wget $url/ST-CMDS-20170001_1-OS.tar.gz
+    tar xf ST-CMDS-20170001_1-OS.tar.gz
+
+    if $remove_archive; then
+        echo "$0: removing $data/ST-CMDS-20170001_1-OS.tar.gz file since --remove-archive option was supplied."
+        rm $data/ST-CMDS-20170001_1-OS.tar.gz
+    fi
+
+    cd "${cwd}" || exit 1;
+    echo "$0: Successfully downloaded and un-tarred $data/ST-CMDS-20170001_1-OS.tar.gz"
+    touch ${data}/.complete
+else
+    echo "$0: Already exists. Skip download."
+fi
+
+exit 0;
diff --git a/egs2/zh_openslr38/asr1/local/data_split.py b/egs2/zh_openslr38/asr1/local/data_split.py
new file mode 100644
index 00000000000..e76b091a8ed
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/local/data_split.py
@@ -0,0 +1,97 @@
+
+"""
+Split data to train, dev, test
+"""
+import sys
+import os
+from collections import defaultdict
+import random
+
+train_size = 0.9
+
+data_dir = sys.argv[1] #ST-CMDS-20170001_1-OS
+
+# create speaker id dictionary
+d = defaultdict(list)
+for fn in os.listdir(data_dir):
+	if not fn.endswith('.wav'): continue
+	# 20170001P00001A0001.wav
+	prefix, s = fn.split('P')
+	try:
+		speaker, s = s.split('A')
+		letter = 'A'
+	except ValueError:
+		speaker, s = s.split('I')
+		letter = 'I'
+	utt, _ = s.split('.')
+	d[speaker+letter].append(utt)
+
+speaker_ids = list(d.keys())
+random.shuffle(speaker_ids)
+
+num_speakers = len(speaker_ids)
+assert num_speakers == 855, "Number of speakers should be 855 in Free ST Chinese Mandarin Corpus."
+
+num_train = int(train_size*num_speakers)
+num_test = int((num_speakers - num_train)/2)
+
+train_speakers = speaker_ids[:num_train]
+dev_speakers = speaker_ids[num_train:-num_test]
+test_speakers = speaker_ids[-num_test:]
+
+print(f"# train: {num_train}, # dev:{num_speakers-num_train-num_test}, # test:{num_test}")
+
+def get_transcription(spk_id, utt_id):
+	text_fn = get_text_filename(spk_id, utt_id)
+	with open(text_fn) as f:
+		lines = f.readlines()
+	assert len(lines) == 1, f"More than one line in transription file:{text_fn}"
+	return lines[0]
+
+def get_text_filename(spk_id, utt_id):
+	return f"{data_dir}/20170001P{spk_id}{utt_id}.txt"
+
+def get_wav_filename(spk_id, utt_id):
+	return f"{data_dir}/20170001P{spk_id}{utt_id}.wav"
+
+def create_files(speakers, directory):
+	text_lines, scp_lines, utt2spk_lines = [], [], []
+	for spk_id in speakers:
+		for utt_id in d[spk_id]:
+			# add spk_id in front to make utt_id unique
+			unique_utt_id = spk_id+utt_id
+
+			transcription = get_transcription(spk_id, utt_id)
+			text_lines.append(f"{unique_utt_id} {transcription}\n")
+
+			wav_file_path = get_wav_filename(spk_id, utt_id)
+			scp_lines.append(f"{unique_utt_id} {wav_file_path}\n")
+
+			utt2spk_lines.append(f"{unique_utt_id} {spk_id}\n")
+
+	# sort
+	text_lines.sort()
+	scp_lines.sort()
+	utt2spk_lines.sort()
+
+	# write to file
+	with open(f"{directory}/text", "w+") as text_file:
+		text_file.writelines(text_lines)
+
+	with open(f"{directory}/wav.scp", "w+") as scp_file:
+		scp_file.writelines(scp_lines)
+	
+	with open(f"{directory}/utt2spk", "w+") as utt2spk_file:
+		utt2spk_file.writelines(utt2spk_lines)
+
+print("Creating files for train...", end="")
+create_files(train_speakers, "data/train")
+print("Done.")
+
+print("Creating files for dev...", end="")
+create_files(dev_speakers, "data/dev")
+print("Done.")
+
+print("Creating files for test...", end="")
+create_files(test_speakers, "data/test")
+print("Done.")
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/local/download_and_untar.sh b/egs2/zh_openslr38/asr1/local/download_and_untar.sh
new file mode 120000
index 00000000000..5e67b127005
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/local/download_and_untar.sh
@@ -0,0 +1 @@
+../../../../egs/aishell/asr1/local/download_and_untar.sh
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/local/path.sh b/egs2/zh_openslr38/asr1/local/path.sh
new file mode 100755
index 00000000000..cd186777d50
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/local/path.sh
@@ -0,0 +1 @@
+MAIN_ROOT=$PWD/../../..
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/path.sh b/egs2/zh_openslr38/asr1/path.sh
new file mode 120000
index 00000000000..c9ac0a75bc6
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/path.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/path.sh
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/pyscripts b/egs2/zh_openslr38/asr1/pyscripts
new file mode 120000
index 00000000000..ac68ad75b60
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/pyscripts
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/pyscripts
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/run.sh b/egs2/zh_openslr38/asr1/run.sh
new file mode 100755
index 00000000000..fb1d737316a
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/run.sh
@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+train_set=train
+valid_set=dev
+test_sets="dev test"
+
+asr_config=conf/train_asr_conformer.yaml
+inference_config=conf/decode_asr_rnn.yaml
+
+lm_config=conf/train_lm_transformer.yaml
+use_lm=true
+use_wordlm=false
+
+# speed perturbation related
+# (train_set will be "${train_set}_sp" if speed_perturb_factors is specified)
+speed_perturb_factors="0.9 1.0 1.1"
+
+./asr.sh                                               \
+    --lang zh                                          \
+    --audio_format wav                                 \
+    --feats_type raw                                   \
+    --token_type char                                  \
+    --use_lm ${use_lm}                                 \
+    --use_word_lm ${use_wordlm}                        \
+    --lm_config "${lm_config}"                         \
+    --asr_config "${asr_config}"                       \
+    --inference_config "${inference_config}"           \
+    --train_set "${train_set}"                         \
+    --valid_set "${valid_set}"                         \
+    --test_sets "${test_sets}"                         \
+    --speed_perturb_factors "${speed_perturb_factors}" \
+    --asr_speech_fold_length 512 \
+    --asr_text_fold_length 150 \
+    --lm_fold_length 150 \
+    --lm_train_text "data/${train_set}/text" "$@"
diff --git a/egs2/zh_openslr38/asr1/run_streaming.sh b/egs2/zh_openslr38/asr1/run_streaming.sh
new file mode 100755
index 00000000000..8149ae6e842
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/run_streaming.sh
@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+train_set=train
+valid_set=dev
+test_sets="dev test"
+
+asr_config=conf/train_asr_streaming_transformer.yaml
+inference_config=conf/decode_asr_streaming.yaml
+
+lm_config=conf/train_lm.yaml
+use_lm=true
+use_wordlm=false
+
+# speed perturbation related
+# (train_set will be "${train_set}_sp" if speed_perturb_factors is specified)
+speed_perturb_factors="0.9 1.0 1.1"
+
+./asr.sh                                               \
+    --use_streaming true                               \
+    --lang zh                                          \
+    --audio_format wav                                 \
+    --feats_type raw                                   \
+    --token_type char                                  \
+    --use_lm ${use_lm}                                 \
+    --use_word_lm ${use_wordlm}                        \
+    --lm_config "${lm_config}"                         \
+    --asr_config "${asr_config}"                       \
+    --inference_config "${inference_config}"           \
+    --train_set "${train_set}"                         \
+    --valid_set "${valid_set}"                         \
+    --test_sets "${test_sets}"                         \
+    --speed_perturb_factors "${speed_perturb_factors}" \
+    --asr_speech_fold_length 512 \
+    --asr_text_fold_length 150 \
+    --lm_fold_length 150 \
+    --lm_train_text "data/${train_set}/text" "$@"
diff --git a/egs2/zh_openslr38/asr1/scripts b/egs2/zh_openslr38/asr1/scripts
new file mode 120000
index 00000000000..b25829705dc
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/scripts
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/scripts
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/steps b/egs2/zh_openslr38/asr1/steps
new file mode 120000
index 00000000000..91f2d234e20
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/steps
@@ -0,0 +1 @@
+../../../tools/kaldi/egs/wsj/s5/steps
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/utils b/egs2/zh_openslr38/asr1/utils
new file mode 120000
index 00000000000..f49247da827
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/utils
@@ -0,0 +1 @@
+../../../tools/kaldi/egs/wsj/s5/utils
\ No newline at end of file

From 6c8cb2954a4bf19ad6a72c4fa5bbb76d4f843bc1 Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-37-72.us-east-2.compute.internal>
Date: Wed, 16 Mar 2022 03:42:15 +0000
Subject: [PATCH 03/22] remove duplicates in test

---
 .../utils/check_train_test_duplicate.py       | 57 +++++++++++++++++++
 egs2/zh_openslr38/asr1/local/data.sh          |  1 +
 2 files changed, 58 insertions(+)
 create mode 100644 egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py

diff --git a/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py b/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
new file mode 100644
index 00000000000..11e5b6fa7da
--- /dev/null
+++ b/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
@@ -0,0 +1,57 @@
+from collections import Counter 
+
+train_file = "data/train/text"
+
+train_lines = []
+with open(train_file) as f:
+    for line in f:
+        if not line: continue
+        train_lines.append(line.split()[1])
+train_lines = set(train_lines)
+
+for test_name in ("test", "dev"):
+    test_file = f"data/{test_name}/text"
+
+    test_lines = []
+    test_uttids = []
+    with open(test_file) as f:
+        for line in f:
+            if not line: continue
+            test_uttids.append(line.split()[0])
+            test_lines.append(line.split()[1])
+
+    count = 0
+    duplicate_uttids = []  # duplicate ids in the test file
+    for t, uttid in zip(test_lines, test_uttids):
+        if t in train_lines:
+            duplicate_uttids.append(uttid)
+            count += 1
+    duplicate_uttids = set(duplicate_uttids)
+    print(count, "duplicates in", test_name)
+
+    if input("continue? [y/n]") == 'y':
+        # remove all instances of duplicate uttids in these files: spk2utt, text, utt2spk, wav.scp
+        with open(f"data/{test_name}/spk2utt", 'r') as f:
+            # replace all uttid with empty string
+            text = f.read()
+            for uttid in duplicate_uttids:
+                text = text.replace(" "+uttid, "")
+            for line in text.split("\n"):
+                if not line: continue
+                if len(line.strip().split(" ")) < 2:
+                    print(f"removing {line} from spk2utt")
+                    text = text.replace(line+'\n', "")
+        with open(f"data/{test_name}/spk2utt", 'w') as f:
+            f.write(text)   
+
+        for name in ("text", "utt2spk", "wav.scp"):
+            with open(f"data/{test_name}/{name}", "r") as f:
+                # remove all lines that contain ids that correspond to duplicate sentences
+                out_lines = []
+                for line in f:
+                    if not line.split()[0] in duplicate_uttids:
+                        out_lines.append(line.strip())
+            with open(f"data/{test_name}/{name}", "w") as f:
+                f.write("\n".join(out_lines))
+    else:
+        print("ok.")
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/local/data.sh b/egs2/zh_openslr38/asr1/local/data.sh
index 002756eb78c..f32eaf42f10 100755
--- a/egs2/zh_openslr38/asr1/local/data.sh
+++ b/egs2/zh_openslr38/asr1/local/data.sh
@@ -72,6 +72,7 @@ mkdir -p $dev_dir
 mkdir -p $test_dir
 
 python3 local/data_split.py ${ST_CMDS}/ST-CMDS-20170001_1-OS
+python3 pyscripts/utils/check_train_test_duplicate.py
 
 for dir in $train_dir $dev_dir $test_dir; do
   utils/utt2spk_to_spk2utt.pl $dir/utt2spk > $dir/spk2utt

From 7fe99ff74d968fc9b9d6c91f403b1d0bb04c2658 Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-37-72.us-east-2.compute.internal>
Date: Wed, 16 Mar 2022 03:50:52 +0000
Subject: [PATCH 04/22] add newline and random seed

---
 egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py | 1 +
 egs2/zh_openslr38/asr1/local/data_split.py                       | 1 +
 2 files changed, 2 insertions(+)

diff --git a/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py b/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
index 11e5b6fa7da..7167f0c408b 100644
--- a/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
+++ b/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
@@ -53,5 +53,6 @@
                         out_lines.append(line.strip())
             with open(f"data/{test_name}/{name}", "w") as f:
                 f.write("\n".join(out_lines))
+                f.write("\n")
     else:
         print("ok.")
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/local/data_split.py b/egs2/zh_openslr38/asr1/local/data_split.py
index e76b091a8ed..60b351f6965 100644
--- a/egs2/zh_openslr38/asr1/local/data_split.py
+++ b/egs2/zh_openslr38/asr1/local/data_split.py
@@ -8,6 +8,7 @@
 import random
 
 train_size = 0.9
+random.seed(1)
 
 data_dir = sys.argv[1] #ST-CMDS-20170001_1-OS
 

From 12277a27c5d2bf6f5a3c1fa23c21e06dbab5d25f Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-37-72.us-east-2.compute.internal>
Date: Wed, 16 Mar 2022 04:40:16 +0000
Subject: [PATCH 05/22] bugfix: move pre-processing order

---
 egs2/zh_openslr38/asr1/local/data.sh | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/egs2/zh_openslr38/asr1/local/data.sh b/egs2/zh_openslr38/asr1/local/data.sh
index f32eaf42f10..15c0341c7eb 100755
--- a/egs2/zh_openslr38/asr1/local/data.sh
+++ b/egs2/zh_openslr38/asr1/local/data.sh
@@ -72,12 +72,13 @@ mkdir -p $dev_dir
 mkdir -p $test_dir
 
 python3 local/data_split.py ${ST_CMDS}/ST-CMDS-20170001_1-OS
-python3 pyscripts/utils/check_train_test_duplicate.py
 
 for dir in $train_dir $dev_dir $test_dir; do
   utils/utt2spk_to_spk2utt.pl $dir/utt2spk > $dir/spk2utt
 done
 
+python3 pyscripts/utils/check_train_test_duplicate.py
+
 # validate formats
 utils/validate_data_dir.sh --no-feats data/train
 utils/validate_data_dir.sh --no-feats data/dev

From f9c8fa0607f093c7a735d0b6d19058fbf906890f Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-18-124.us-east-2.compute.internal>
Date: Wed, 16 Mar 2022 05:02:57 +0000
Subject: [PATCH 06/22] remove user confirmation in check duplicate script

---
 .../utils/check_train_test_duplicate.py       | 52 +++++++++----------
 1 file changed, 26 insertions(+), 26 deletions(-)

diff --git a/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py b/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
index 7167f0c408b..cb1406bee54 100644
--- a/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
+++ b/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
@@ -29,30 +29,30 @@
     duplicate_uttids = set(duplicate_uttids)
     print(count, "duplicates in", test_name)
 
-    if input("continue? [y/n]") == 'y':
-        # remove all instances of duplicate uttids in these files: spk2utt, text, utt2spk, wav.scp
-        with open(f"data/{test_name}/spk2utt", 'r') as f:
-            # replace all uttid with empty string
-            text = f.read()
-            for uttid in duplicate_uttids:
-                text = text.replace(" "+uttid, "")
-            for line in text.split("\n"):
-                if not line: continue
-                if len(line.strip().split(" ")) < 2:
-                    print(f"removing {line} from spk2utt")
-                    text = text.replace(line+'\n', "")
-        with open(f"data/{test_name}/spk2utt", 'w') as f:
-            f.write(text)   
+    # if input("continue? [y/n]") == 'y':
+    # remove all instances of duplicate uttids in these files: spk2utt, text, utt2spk, wav.scp
+    with open(f"data/{test_name}/spk2utt", 'r') as f:
+        # replace all uttid with empty string
+        text = f.read()
+        for uttid in duplicate_uttids:
+            text = text.replace(" "+uttid, "")
+        for line in text.split("\n"):
+            if not line: continue
+            if len(line.strip().split(" ")) < 2:
+                print(f"removing {line} from spk2utt")
+                text = text.replace(line+'\n', "")
+    with open(f"data/{test_name}/spk2utt", 'w') as f:
+        f.write(text)   
 
-        for name in ("text", "utt2spk", "wav.scp"):
-            with open(f"data/{test_name}/{name}", "r") as f:
-                # remove all lines that contain ids that correspond to duplicate sentences
-                out_lines = []
-                for line in f:
-                    if not line.split()[0] in duplicate_uttids:
-                        out_lines.append(line.strip())
-            with open(f"data/{test_name}/{name}", "w") as f:
-                f.write("\n".join(out_lines))
-                f.write("\n")
-    else:
-        print("ok.")
\ No newline at end of file
+    for name in ("text", "utt2spk", "wav.scp"):
+        with open(f"data/{test_name}/{name}", "r") as f:
+            # remove all lines that contain ids that correspond to duplicate sentences
+            out_lines = []
+            for line in f:
+                if not line.split()[0] in duplicate_uttids:
+                    out_lines.append(line.strip())
+        with open(f"data/{test_name}/{name}", "w") as f:
+            f.write("\n".join(out_lines))
+            f.write("\n")
+    # else:
+    #     print("ok.")
\ No newline at end of file

From dd98c3a97e8e92799e24d8c8fc63147ebeab3568 Mon Sep 17 00:00:00 2001
From: Shinji Watanabe <sw005320@gmail.com>
Date: Thu, 17 Mar 2022 07:53:53 -0400
Subject: [PATCH 07/22] Fix n-gram decoding

---
 espnet2/bin/asr_inference.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/espnet2/bin/asr_inference.py b/espnet2/bin/asr_inference.py
index fbb156d8229..d5a7932b0e6 100755
--- a/espnet2/bin/asr_inference.py
+++ b/espnet2/bin/asr_inference.py
@@ -136,6 +136,7 @@ def __init__(
                 decoder=1.0 - ctc_weight,
                 ctc=ctc_weight,
                 lm=lm_weight,
+                ngram=ngram_weight,
                 length_bonus=penalty,
             )
             beam_search = BeamSearch(

From 37865898d14e2fa22885a010371f03ddf56977de Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-18-124.us-east-2.compute.internal>
Date: Fri, 18 Mar 2022 21:56:26 +0000
Subject: [PATCH 08/22] add scripts for hubert self superised learning

---
 .../asr1/conf/train_asr_conformer_sslr.yaml   | 89 +++++++++++++++++++
 egs2/zh_openslr38/asr1/run_sslr.sh            | 43 +++++++++
 2 files changed, 132 insertions(+)
 create mode 100644 egs2/zh_openslr38/asr1/conf/train_asr_conformer_sslr.yaml
 create mode 100755 egs2/zh_openslr38/asr1/run_sslr.sh

diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_conformer_sslr.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_conformer_sslr.yaml
new file mode 100644
index 00000000000..03a410cded5
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/conf/train_asr_conformer_sslr.yaml
@@ -0,0 +1,89 @@
+# network architecture
+# encoder related
+encoder: conformer
+encoder_conf:
+    output_size: 256    # dimension of attention
+    attention_heads: 4
+    linear_units: 2048  # the number of units of position-wise feed forward
+    num_blocks: 12      # the number of encoder blocks
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    attention_dropout_rate: 0.0
+    input_layer: conv2d # encoder architecture type
+    normalize_before: true
+    pos_enc_layer_type: rel_pos
+    selfattention_layer_type: rel_selfattn
+    activation_type: swish
+    macaron_style: true
+    use_cnn_module: true
+    cnn_module_kernel: 15
+
+# decoder related
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 2048
+    num_blocks: 6
+    dropout_rate: 0.1
+    positional_dropout_rate: 0.1
+    self_attention_dropout_rate: 0.0
+    src_attention_dropout_rate: 0.0
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1     # label smoothing option
+    length_normalized_loss: false
+    extract_feats_in_collect_stats: false   # Note: "False" means during collect stats (stage 10), generating dummy stats files rather than extract_feats by forward frontend.
+
+# minibatch related
+batch_type: numel
+batch_bins: 4000000
+
+# optimization related
+accum_grad: 4
+grad_clip: 5
+max_epoch: 40
+val_scheduler_criterion:
+    - valid
+    - acc
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 10
+
+optim: adam
+optim_conf:
+   lr: 0.0005
+scheduler: warmuplr
+scheduler_conf:
+   warmup_steps: 30000
+
+frontend: s3prl
+frontend_conf:
+   frontend_conf:
+      upstream: hubert_large_ll60k  # Note: If the upstream is changed, please change the input_size in the preencoder.
+   download_dir: ./hub
+   multilayer_feature: True
+   
+preencoder: linear
+preencoder_conf:
+   input_size: 1024  # Note: If the upstream is changed, please change this value accordingly.
+   output_size: 80
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 40
+    num_time_mask: 2
diff --git a/egs2/zh_openslr38/asr1/run_sslr.sh b/egs2/zh_openslr38/asr1/run_sslr.sh
new file mode 100755
index 00000000000..1e391b9379e
--- /dev/null
+++ b/egs2/zh_openslr38/asr1/run_sslr.sh
@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+train_set=train
+valid_set=dev
+test_sets="dev test"
+
+asr_config=conf/train_asr_conformer_sslr.yaml
+inference_config=conf/decode_asr_rnn.yaml
+
+lm_config=conf/train_lm_transformer.yaml
+use_lm=true
+use_wordlm=false
+
+# speed perturbation related
+# (train_set will be "${train_set}_sp" if speed_perturb_factors is specified)
+speed_perturb_factors="0.9 1.0 1.1"
+
+./asr.sh                                               \
+    --lang zh                                          \
+    --audio_format wav                                 \
+    --feats_type raw                                   \
+    --token_type char                                  \
+    --use_lm ${use_lm}                                 \
+    --use_word_lm ${use_wordlm}                        \
+    --lm_config "${lm_config}"                         \
+    --asr_config "${asr_config}"                       \
+    --inference_config "${inference_config}"           \
+    --train_set "${train_set}"                         \
+    --valid_set "${valid_set}"                         \
+    --test_sets "${test_sets}"                         \
+    --speed_perturb_factors "${speed_perturb_factors}" \
+    --asr_speech_fold_length 512 \
+    --asr_text_fold_length 150 \
+    --lm_fold_length 150 \
+    --lm_train_text "data/${train_set}/text" "$@" \
+    --feats_normalize uttmvn \
+    --nj 1 \
+    --stage 10

From ef6dcc76a9d1e0f4cb08708bd05a97b40bce5c76 Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-37-72.us-east-2.compute.internal>
Date: Sun, 20 Mar 2022 19:04:10 +0000
Subject: [PATCH 09/22] update README

---
 egs2/zh_openslr38/asr1/README.md | 180 ++-----------------------------
 1 file changed, 11 insertions(+), 169 deletions(-)

diff --git a/egs2/zh_openslr38/asr1/README.md b/egs2/zh_openslr38/asr1/README.md
index 985604c1a43..57b9e7f2aa2 100644
--- a/egs2/zh_openslr38/asr1/README.md
+++ b/egs2/zh_openslr38/asr1/README.md
@@ -1,179 +1,21 @@
-# Streaming Conformer + specaug + speed perturbation: feats=raw, n_fft=512, hop_length=128
-## Environments
-- date: `Mon Aug 23 16:31:48 CST 2021`
-- python version: `3.7.9 (default, Aug 31 2020, 12:42:55)  [GCC 7.3.0]`
-- espnet version: `espnet 0.9.9`
-- pytorch version: `pytorch 1.5.0`
-- Git hash: `b94d07028099a80c9c690341981ae7d550b5ca24`
-  - Commit date: `Mon Aug 23 00:47:47 2021 +0800`
-
-## With Transformer LM
-- Model link: (wait for upload)
-- ASR config: [./conf/train_asr_streaming_cpnformer.yaml](./conf/train_asr_streaming_conformer.yaml)
-- LM config: [./conf/tuning/train_lm_transformer.yaml](./conf/tuning/train_lm_transformer.yaml)
-
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_streaming_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.0|5.8|0.3|0.3|6.3|42.2|
-|decode_asr_streaming_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|92.9|6.7|0.5|0.7|7.8|46.2|
-# Streaming Transformer + speed perturbation: feats=raw, n_fft=512, hop_length=128
-## Environments
-- date: `Tue Aug 17 01:20:32 CST 2021`
-- python version: `3.7.9 (default, Aug 31 2020, 12:42:55)  [GCC 7.3.0]`
-- espnet version: `espnet 0.9.9`
-- pytorch version: `pytorch 1.5.0`
-- Git hash: `6f5f848e0a9bfca1b73393779233bde34add3df1`
-  - Commit date: `Mon Aug 16 21:50:08 2021 +0800`
-
-## With Transformer LM
-- Model link: (wait for upload)
-- ASR config: [./conf/train_asr_streaming_transformer.yaml](./conf/train_asr_streaming_transformer.yaml)
-- LM config: [./conf/tuning/train_lm_transformer.yaml](./conf/tuning/train_lm_transformer.yaml)
-
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_streaming_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|93.6|6.2|0.1|0.5|6.8|46.8|
-|decode_asr_streaming_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.0|6.7|0.2|0.8|7.8|50.7|
-
-# Conformer + specaug + speed perturbation: feats=raw, n_fft=512, hop_length=128
-## Environments
-- date: `Fri Oct 16 11:10:17 JST 2020`
-- python version: `3.7.3 (default, Mar 27 2019, 22:11:17)  [GCC 7.3.0]`
-- espnet version: `espnet 0.9.0`
-- pytorch version: `pytorch 1.6.0`
-- Git hash: `20b0c89369d9dd3e05780b65fdd00a9b4f4891e5`
-  - Commit date: `Mon Oct 12 09:28:20 2020 -0400`
-
-## With Transformer LM
-- Model link: https://zenodo.org/record/4105763#.X40xe2j7QUE
-- ASR config: [./conf/tuning/train_asr_conformer.yaml](./conf/tuning/train_asr_conformer.yaml)
-- LM config: [./conf/tuning/train_lm_transformer.yaml](./conf/tuning/train_lm_transformer.yaml)
-
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_rnn_lm_lm_train_lm_transformer_char_batch_bins2000000_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|95.7|4.2|0.1|0.1|4.4|33.7|
-|decode_asr_rnn_lm_lm_train_lm_transformer_char_batch_bins2000000_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|95.4|4.5|0.1|0.1|4.7|35.0|
-
-## With RNN LM
-- ASR config: [./conf/tuning/train_asr_conformer.yaml](./conf/tuning/train_asr_conformer.yaml)
-- LM config: [./conf/tuning/train_lm_rnn2.yaml](./conf/tuning/train_lm_rnn2.yaml)
-
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|95.5|4.4|0.1|0.1|4.6|35.2|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|95.2|4.7|0.1|0.1|4.9|36.5|
-
-## Without LM
-- ASR config: [./conf/tuning/train_asr_conformer.yaml](./conf/tuning/train_asr_conformer.yaml)
-
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_rnn_asr_model_valid.acc.ave/dev|14326|205341|95.6|4.3|0.1|0.1|4.5|35.0|
-|decode_asr_rnn_asr_model_valid.acc.ave/test|7176|104765|95.2|4.7|0.1|0.1|4.9|36.7|
-
-# Transformer + speed perturbation: feats=raw with same LM with the privious setting
-
-I compared between `n_fft=512, hop_length=128`, `n_fft=400, hop_length=160`,  and `n_fft=512, hop_length=256`
-with searching the best `batch_bins` to get the suitable configuration for each hop_length.
+# Corpus
+Free ST Chinese Mandarin Corpus: a free Mandarin Chinese corpus collected by Surfingtech (www.surfing.ai). The dataset contains 102600 utterances from 855 speakers, for a total of 109.73 hours of speech. 
 
-## Environments
-- date: `Fri Oct 16 11:10:17 JST 2020`
-- python version: `3.7.3 (default, Mar 27 2019, 22:11:17)  [GCC 7.3.0]`
-- espnet version: `espnet 0.9.0`
-- pytorch version: `pytorch 1.6.0`
-- Git hash: `20b0c89369d9dd3e05780b65fdd00a9b4f4891e5`
-  - Commit date: `Mon Oct 12 09:28:20 2020 -0400`
-
-## n_fft=512, hop_length=128
-asr_train_asr_transformer2_raw_char_batch_typenumel_batch_bins8500000_optim_conflr0.0005_scheduler_confwarmup_steps30000_sp
-
-- ASR config: [./conf/tuning/train_asr_transformer3.yaml](./conf/tuning/train_asr_transformer3.yaml)
-- LM config: [./conf/tuning/train_lm_rnn2.yaml](./conf/tuning/train_lm_rnn2.yaml)
-
-
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.2|5.7|0.1|0.1|5.9|42.6|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.7|6.1|0.2|0.1|6.4|45.0|
-
-
-## n_fft=400, hop_length=160
-asr_train_asr_transformer2_raw_char_frontend_confn_fft400_frontend_confhop_length160_batch_typenumel_batch_bins6500000_optim_conflr0.0005_scheduler_confwarmup_steps30000_sp
-
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.1|5.7|0.1|0.1|6.0|43.0|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.5|6.3|0.2|0.1|6.6|45.4|
-
-## n_fft=512, hop_length=256
-asr_train_asr_transformer2_raw_char_frontend_confn_fft512_frontend_confhop_length256_batch_typenumel_batch_bins6000000_sp
-
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.0|5.9|0.1|0.1|6.1|43.5|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.3|6.5|0.2|0.1|6.8|45.8|
-
-
-# Transformer + speed perturbation: feats=fbank_pitch, RNN-LM
-Compatible setting with espnet1 to reproduce the previou result
-
-## Environments
-- date: `Fri Oct 16 11:10:17 JST 2020`
-- python version: `3.7.3 (default, Mar 27 2019, 22:11:17)  [GCC 7.3.0]`
-- espnet version: `espnet 0.9.0`
-- pytorch version: `pytorch 1.6.0`
-- Git hash: `20b0c89369d9dd3e05780b65fdd00a9b4f4891e5`
-  - Commit date: `Mon Oct 12 09:28:20 2020 -0400`
-
-- ASR config: [./conf/tuning/train_asr_transformer2.yaml](./conf/tuning/train_asr_transformer2.yaml)
-- LM config: [./conf/tuning/train_lm_rnn2.yaml](./conf/tuning/train_lm_rnn2.yaml)
+Since all speakers have 120 utterances, we manually divide the data into train, dev, and test split with a ratio of 90-5-5 using speaker IDs, resulting in 769, 43, and 43 speakers in our train, dev, test split respectively. Utterances with the same speaker ID are kept in the same split.
 
-## asr_train_asr_transformer2_fbank_pitch_char_sp
-### CER
+The original dataset contains duplicates sentences with the same transcript, but are spoken by different speakers. Although the waveforms are different for these duplicates, we still remove sentences in the test and development set that have duplicate transcripts in the training set, in order to eliminate any effect of training data leakage.
 
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/dev|14326|205341|94.0|5.9|0.1|0.1|6.1|43.4|
-|decode_asr_rnn_lm_lm_train_lm_char_valid.loss.ave_asr_model_valid.acc.ave/test|7176|104765|93.4|6.4|0.2|0.1|6.7|45.9|
+Link: https://openslr.org/38
 
-# The first result
+# Results
 ## Environments
-- date: `Sun Feb  2 02:03:55 CST 2020`
-- python version: `3.7.3 (default, Mar 27 2019, 22:11:17)  [GCC 7.3.0]`
-- espnet version: `espnet 0.6.0`
-- pytorch version: `pytorch 1.1.0`
-- Git hash: `e0fd073a70bcded6a0e6a3587630410a994ccdb8` (+ fixing https://github.com/espnet/espnet/pull/1533)
-  - Commit date: `Sat Jan 11 06:09:24 2020 +0900`
+- python version: `3.9.10 | packaged by conda-forge | (main, Feb  1 2022, 21:24:11)  [GCC 9.4.0]`
+- espnet version: `espnet 0.10.7a1`
+- pytorch version: `pytorch 1.10.1`
 
-## asr_train_asr_rnn_new_fbank_pitch_char
 ### CER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
-|decode_devdecode_asr_rnn_lm_valid.loss.best_asr_model_valid.acc.best|14326|205341|92.6|7.2|0.2|0.1|7.5|49.6|
-|decode_testdecode_asr_rnn_lm_valid.loss.best_asr_model_valid.acc.best|7176|104765|91.6|8.2|0.3|0.2|8.6|53.4|
-
-## asr_train_asr_transformer_lr0.002_fbank_pitch_char
-### CER
-
-|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
-|---|---|---|---|---|---|---|---|---|
-|decode_dev_decode_asr_rnn_lm_train_lm_char_valid.loss.best_asr_model_valid.acc.best|14326|205341|93.3|6.5|0.2|0.1|6.8|45.6|
-|decode_test_decode_asr_rnn_lm_train_lm_char_valid.loss.best_asr_model_valid.acc.best|7176|104765|92.7|7.1|0.3|0.1|7.4|47.6|
-
+|decode_asr_rnn_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/dev|4322|46490|91.0|8.4|0.5|0.2|9.2|51.5|
+|decode_asr_rnn_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/test|4167|45803|91.1|8.5|0.5|0.2|9.1|52.2|

From a8bfd94eab04a184791bc0bc1c0a441882f4d673 Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-37-72.us-east-2.compute.internal>
Date: Sun, 20 Mar 2022 19:05:34 +0000
Subject: [PATCH 10/22] update README

---
 egs2/zh_openslr38/asr1/README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/egs2/zh_openslr38/asr1/README.md b/egs2/zh_openslr38/asr1/README.md
index 57b9e7f2aa2..40fa2bc49d6 100644
--- a/egs2/zh_openslr38/asr1/README.md
+++ b/egs2/zh_openslr38/asr1/README.md
@@ -1,5 +1,5 @@
 # Corpus
-Free ST Chinese Mandarin Corpus: a free Mandarin Chinese corpus collected by Surfingtech (www.surfing.ai). The dataset contains 102600 utterances from 855 speakers, for a total of 109.73 hours of speech. 
+**Free ST Chinese Mandarin Corpus**: a free Mandarin Chinese corpus collected by Surfingtech (www.surfing.ai). The dataset contains 102600 utterances from 855 speakers, for a total of 109.73 hours of speech. 
 
 Since all speakers have 120 utterances, we manually divide the data into train, dev, and test split with a ratio of 90-5-5 using speaker IDs, resulting in 769, 43, and 43 speakers in our train, dev, test split respectively. Utterances with the same speaker ID are kept in the same split.
 

From 20e0b8524610fbe83068f1e56d123af63c2a7b6d Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-18-124.us-east-2.compute.internal>
Date: Mon, 21 Mar 2022 20:35:14 +0000
Subject: [PATCH 11/22] clean up and reformat files for PR

---
 .../{decode_asr_rnn.yaml => decode_asr.yaml}  |  0
 .../asr1/conf/decode_asr_simu_streaming.yaml  |  7 --
 .../asr1/conf/decode_asr_streaming.yaml       | 11 ---
 .../asr1/conf/decode_asr_transformer.yaml     |  6 --
 .../conf/decode_asr_transformer_ngram.yaml    |  7 --
 ...rain_asr_conformer.yaml => train_asr.yaml} |  2 +-
 .../asr1/conf/train_asr_conformer.yaml        |  1 -
 .../zh_openslr38/asr1/conf/train_asr_rnn.yaml | 55 -----------
 .../conf/train_asr_streaming_conformer.yaml   | 95 -------------------
 .../conf/train_asr_streaming_transformer.yaml | 76 ---------------
 .../asr1/conf/train_asr_transformer.yaml      |  1 -
 egs2/zh_openslr38/asr1/conf/train_lm.yaml     | 30 ++++--
 .../asr1/conf/train_lm_transformer.yaml       |  1 -
 .../train_asr_sslr.yaml}                      |  0
 .../conf/tuning/train_asr_transformer.yaml    | 66 -------------
 .../conf/tuning/train_asr_transformer2.yaml   | 54 -----------
 .../conf/tuning/train_asr_transformer3.yaml   | 70 --------------
 .../tuning/train_asr_transformer_lr0.002.yaml | 66 -------------
 .../asr1/conf/tuning/train_lm_rnn.yaml        | 19 ----
 .../asr1/conf/tuning/train_lm_rnn2.yaml       | 19 ----
 .../conf/tuning/train_lm_transformer.yaml     | 29 ------
 .../asr1/local}/check_train_test_duplicate.py |  0
 egs2/zh_openslr38/asr1/local/data.sh          |  2 +-
 .../asr1/local/download_and_untar.sh          |  1 -
 egs2/zh_openslr38/asr1/run.sh                 |  6 +-
 egs2/zh_openslr38/asr1/run_sslr.sh            |  9 +-
 egs2/zh_openslr38/asr1/run_streaming.sh       | 41 --------
 27 files changed, 30 insertions(+), 644 deletions(-)
 rename egs2/zh_openslr38/asr1/conf/{decode_asr_rnn.yaml => decode_asr.yaml} (100%)
 delete mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_simu_streaming.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_streaming.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_transformer.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/decode_asr_transformer_ngram.yaml
 rename egs2/zh_openslr38/asr1/conf/{tuning/train_asr_conformer.yaml => train_asr.yaml} (98%)
 delete mode 120000 egs2/zh_openslr38/asr1/conf/train_asr_conformer.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/train_asr_rnn.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/train_asr_streaming_conformer.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/train_asr_streaming_transformer.yaml
 delete mode 120000 egs2/zh_openslr38/asr1/conf/train_asr_transformer.yaml
 delete mode 120000 egs2/zh_openslr38/asr1/conf/train_lm_transformer.yaml
 rename egs2/zh_openslr38/asr1/conf/{train_asr_conformer_sslr.yaml => tuning/train_asr_sslr.yaml} (100%)
 delete mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer2.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer3.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer_lr0.002.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn2.yaml
 delete mode 100644 egs2/zh_openslr38/asr1/conf/tuning/train_lm_transformer.yaml
 rename egs2/{TEMPLATE/asr1/pyscripts/utils => zh_openslr38/asr1/local}/check_train_test_duplicate.py (100%)
 delete mode 120000 egs2/zh_openslr38/asr1/local/download_and_untar.sh
 delete mode 100755 egs2/zh_openslr38/asr1/run_streaming.sh

diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_rnn.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr.yaml
similarity index 100%
rename from egs2/zh_openslr38/asr1/conf/decode_asr_rnn.yaml
rename to egs2/zh_openslr38/asr1/conf/decode_asr.yaml
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_simu_streaming.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_simu_streaming.yaml
deleted file mode 100644
index f79545ddb2f..00000000000
--- a/egs2/zh_openslr38/asr1/conf/decode_asr_simu_streaming.yaml
+++ /dev/null
@@ -1,7 +0,0 @@
-beam_size: 20
-penalty: 0.0
-maxlenratio: 0.0
-minlenratio: 0.0
-ctc_weight: 0.6
-lm_weight: 0.3
-streaming: True
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_streaming.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_streaming.yaml
deleted file mode 100644
index fc2b3b1af5d..00000000000
--- a/egs2/zh_openslr38/asr1/conf/decode_asr_streaming.yaml
+++ /dev/null
@@ -1,11 +0,0 @@
-beam_size: 10
-penalty: 0.0
-maxlenratio: 0.0
-minlenratio: 0.0
-ctc_weight: 0.6
-lm_weight: 0.3
-sim_chunk_length: 512
-disable_repetition_detection: true
-decoder_text_length_limit: 0
-encoded_feat_length_limit: 0
-#streaming: True
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_transformer.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_transformer.yaml
deleted file mode 100644
index a147fa79dc8..00000000000
--- a/egs2/zh_openslr38/asr1/conf/decode_asr_transformer.yaml
+++ /dev/null
@@ -1,6 +0,0 @@
-beam_size: 10
-penalty: 0.0
-maxlenratio: 0.0
-minlenratio: 0.0
-ctc_weight: 0.5
-lm_weight: 0.7
diff --git a/egs2/zh_openslr38/asr1/conf/decode_asr_transformer_ngram.yaml b/egs2/zh_openslr38/asr1/conf/decode_asr_transformer_ngram.yaml
deleted file mode 100644
index 5e3e8e049d4..00000000000
--- a/egs2/zh_openslr38/asr1/conf/decode_asr_transformer_ngram.yaml
+++ /dev/null
@@ -1,7 +0,0 @@
-beam_size: 10
-penalty: 0.0
-maxlenratio: 0.0
-minlenratio: 0.0
-ctc_weight: 0.5
-lm_weight: 0.7
-ngram_weight: 0.9
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_conformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr.yaml
similarity index 98%
rename from egs2/zh_openslr38/asr1/conf/tuning/train_asr_conformer.yaml
rename to egs2/zh_openslr38/asr1/conf/train_asr.yaml
index ab51681722f..98588892b1c 100644
--- a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_conformer.yaml
+++ b/egs2/zh_openslr38/asr1/conf/train_asr.yaml
@@ -73,4 +73,4 @@ specaug_conf:
     time_mask_width_range:
     - 0
     - 40
-    num_time_mask: 2
+    num_time_mask: 2
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_conformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_conformer.yaml
deleted file mode 120000
index 10fea4f0971..00000000000
--- a/egs2/zh_openslr38/asr1/conf/train_asr_conformer.yaml
+++ /dev/null
@@ -1 +0,0 @@
-tuning/train_asr_conformer.yaml
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_rnn.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_rnn.yaml
deleted file mode 100644
index eefe1d516fd..00000000000
--- a/egs2/zh_openslr38/asr1/conf/train_asr_rnn.yaml
+++ /dev/null
@@ -1,55 +0,0 @@
-# network architecture
-# encoder related
-encoder: vgg_rnn
-encoder_conf:
-    rnn_type: lstm
-    bidirectional: True
-    use_projection: False
-    num_layers: 3
-    hidden_size: 1024
-    output_size: 1024
-    dropout: 0.0
-    in_channel: 1
-# decoder related
-decoder: rnn
-decoder_conf:
-    rnn_type: lstm
-    num_layers: 2
-    hidden_size: 1024
-    sampling_probability: 0.0   # scheduled sampling option
-    dropout: 0.0
-    att_conf:
-        adim: 1024
-
-# hybrid CTC/attention
-model_conf:
-    ctc_weight: 0.5
-    lsm_weight: 0.0     # label smoothing option
-
-# minibatch related
-batch_type: folded
-batch_size: 30
-
-# optimization related
-init: chainer
-max_epoch: 10
-optim: adadelta
-optim_conf:
-    lr: 1.0
-    rho: 0.95
-    eps: 1.0e-08
-    weight_decay: 0
-patience: 3
-val_scheduler_criterion:
-    - valid
-    - acc
-best_model_criterion:
--   - valid
-    - acc
-    - max
-keep_nbest_models: 1
-scheduler: reducelronplateau
-scheduler_conf:
-    mode: max
-    factor: 0.5
-    patience: 1
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_streaming_conformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_streaming_conformer.yaml
deleted file mode 100644
index 8e4fa086555..00000000000
--- a/egs2/zh_openslr38/asr1/conf/train_asr_streaming_conformer.yaml
+++ /dev/null
@@ -1,95 +0,0 @@
-# network architecture
-# encoder related
-encoder: contextual_block_conformer    # contextual_block_conformer is the core of streaming conformer
-encoder_conf:
-    output_size: 256    # dimension of attention
-    attention_heads: 4
-    linear_units: 2048  # the number of units of position-wise feed forward
-    num_blocks: 12      # the number of encoder blocks
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    attention_dropout_rate: 0.0
-    input_layer: conv2d # encoder architecture type
-    normalize_before: true
-    activation_type: swish
-    macaron_style: true
-    use_cnn_module: true
-    cnn_module_kernel: 15
-    block_size: 40      # streaming configuration
-    hop_size: 16        # streaming configuration
-    look_ahead: 16      # streaming configuration
-    init_average: true  # streaming configuration
-    ctx_pos_enc: true   # streaming configuration
-
-# decoder related
-decoder: transformer
-decoder_conf:
-    attention_heads: 4
-    linear_units: 2048
-    num_blocks: 6
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    self_attention_dropout_rate: 0.0
-    src_attention_dropout_rate: 0.0
-
-# hybrid CTC/attention
-model_conf:
-    ctc_weight: 0.3
-    lsm_weight: 0.1     # label smoothing option
-    length_normalized_loss: false
-
-# minibatch related
-batch_type: folded
-batch_size: 128
-#valid_batch_size: 1
-# optimization related
-accum_grad: 1
-grad_clip: 5
-patience: 3
-max_epoch: 50
-val_scheduler_criterion:
-    - valid
-    - acc
-best_model_criterion:
--   - valid
-    - acc
-    - max
-early_stopping_criterion:
-    - valid
-    - cer_ctc
-    - min
-keep_nbest_models: 10
-
-# NoamLR is deprecated. Use WarmupLR.
-# The following is equivalent setting for NoamLR:
-#
-#    optim: adam
-#    optim_conf:
-#        lr: 10.
-#    scheduler: noamlr
-#    scheduler_conf:
-#        model_size: 256
-#        warmup_steps: 25000
-#
-optim: adam
-optim_conf:
-    lr: 0.0005
-scheduler: warmuplr     # pytorch v1.1.0+ required
-scheduler_conf:
-    warmup_steps: 30000
-num_att_plot: 0
-specaug: specaug
-specaug_conf:
-    apply_time_warp: true
-    time_warp_window: 5
-    time_warp_mode: bicubic
-    apply_freq_mask: true
-    freq_mask_width_range:
-    - 0
-    - 30
-    num_freq_mask: 2
-    apply_time_mask: true
-    time_mask_width_range:
-    - 0
-    - 40
-    num_time_mask: 2
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_streaming_transformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_streaming_transformer.yaml
deleted file mode 100644
index 2abb8749e8d..00000000000
--- a/egs2/zh_openslr38/asr1/conf/train_asr_streaming_transformer.yaml
+++ /dev/null
@@ -1,76 +0,0 @@
-# network architecture
-# encoder related
-encoder: contextual_block_transformer
-encoder_conf:
-    output_size: 256    # dimension of attention
-    attention_heads: 4
-    linear_units: 2048  # the number of units of position-wise feed forward
-    num_blocks: 12      # the number of encoder blocks
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    attention_dropout_rate: 0.0
-    input_layer: conv2d # encoder architecture type
-    normalize_before: true
-    block_size: 40
-    hop_size: 16
-    look_ahead: 16
-    init_average: true
-    ctx_pos_enc: true
-
-# decoder related
-decoder: transformer
-decoder_conf:
-    attention_heads: 4
-    linear_units: 2048
-    num_blocks: 6
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    self_attention_dropout_rate: 0.0
-    src_attention_dropout_rate: 0.0
-
-# hybrid CTC/attention
-model_conf:
-    ctc_weight: 0.3
-    lsm_weight: 0.1     # label smoothing option
-    length_normalized_loss: false
-
-# minibatch related
-batch_type: folded
-batch_size: 128
-#valid_batch_size: 1
-# optimization related
-accum_grad: 1
-grad_clip: 5
-patience: 3
-max_epoch: 20
-val_scheduler_criterion:
-    - valid
-    - acc
-best_model_criterion:
--   - valid
-    - acc
-    - max
-early_stopping_criterion:
-    - valid
-    - cer_ctc
-    - min
-keep_nbest_models: 10
-
-# NoamLR is deprecated. Use WarmupLR.
-# The following is equivalent setting for NoamLR:
-#
-#    optim: adam
-#    optim_conf:
-#        lr: 10.
-#    scheduler: noamlr
-#    scheduler_conf:
-#        model_size: 256
-#        warmup_steps: 25000
-#
-optim: adam
-optim_conf:
-    lr: 0.001
-scheduler: warmuplr     # pytorch v1.1.0+ required
-scheduler_conf:
-    warmup_steps: 25000
-num_att_plot: 0
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_transformer.yaml b/egs2/zh_openslr38/asr1/conf/train_asr_transformer.yaml
deleted file mode 120000
index b9175b1a7fa..00000000000
--- a/egs2/zh_openslr38/asr1/conf/train_asr_transformer.yaml
+++ /dev/null
@@ -1 +0,0 @@
-tuning/train_asr_transformer_lr0.002.yaml
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/conf/train_lm.yaml b/egs2/zh_openslr38/asr1/conf/train_lm.yaml
index 26ea1e4f96e..6f12611bf06 100644
--- a/egs2/zh_openslr38/asr1/conf/train_lm.yaml
+++ b/egs2/zh_openslr38/asr1/conf/train_lm.yaml
@@ -1,19 +1,29 @@
-# rnnlm related
-lm: seq_rnn
+lm: transformer
 lm_conf:
-    unit: 650
-    nlayers: 2
+    pos_enc: null
+    embed_unit: 128
+    att_unit: 512
+    head: 8
+    unit: 2048
+    layer: 16
+    dropout_rate: 0.1
 
 # optimization related
 grad_clip: 5.0
-batch_type: folded
-batch_size: 64  # batch size in LM training
-max_epoch: 20   # if the data size is large, we can reduce this
-patience: 3
-optim: sgd
+batch_type: numel
+batch_bins: 2000000
+accum_grad: 1
+max_epoch: 15  # 15epoch is enougth
+
+optim: adam
+optim_conf:
+   lr: 0.001
+scheduler: warmuplr
+scheduler_conf:
+   warmup_steps: 25000
 
 best_model_criterion:
 -   - valid
     - loss
     - min
-keep_nbest_models: 1
+keep_nbest_models: 10  # 10 is good.
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/conf/train_lm_transformer.yaml b/egs2/zh_openslr38/asr1/conf/train_lm_transformer.yaml
deleted file mode 120000
index 72276561c41..00000000000
--- a/egs2/zh_openslr38/asr1/conf/train_lm_transformer.yaml
+++ /dev/null
@@ -1 +0,0 @@
-tuning/train_lm_transformer.yaml
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/conf/train_asr_conformer_sslr.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_sslr.yaml
similarity index 100%
rename from egs2/zh_openslr38/asr1/conf/train_asr_conformer_sslr.yaml
rename to egs2/zh_openslr38/asr1/conf/tuning/train_asr_sslr.yaml
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer.yaml
deleted file mode 100644
index 1e34e982978..00000000000
--- a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer.yaml
+++ /dev/null
@@ -1,66 +0,0 @@
-# network architecture
-# encoder related
-encoder: transformer
-encoder_conf:
-    output_size: 256    # dimension of attention
-    attention_heads: 4
-    linear_units: 2048  # the number of units of position-wise feed forward
-    num_blocks: 12      # the number of encoder blocks
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    attention_dropout_rate: 0.0
-    input_layer: conv2d # encoder architecture type
-    normalize_before: true
-
-# decoder related
-decoder: transformer
-decoder_conf:
-    attention_heads: 4
-    linear_units: 2048
-    num_blocks: 6
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    self_attention_dropout_rate: 0.0
-    src_attention_dropout_rate: 0.0
-
-# hybrid CTC/attention
-model_conf:
-    ctc_weight: 0.3
-    lsm_weight: 0.1     # label smoothing option
-    length_normalized_loss: false
-
-# minibatch related
-batch_type: folded
-batch_size: 28
-
-# optimization related
-accum_grad: 2
-grad_clip: 5
-patience: 3
-max_epoch: 20
-val_scheduler_criterion:
-    - valid
-    - acc
-best_model_criterion:
--   - valid
-    - acc
-    - max
-keep_nbest_models: 10
-
-# NoamLR is deprecated. Use WarmupLR.
-# The following is equivalent setting for NoamLR:
-#
-#    optim: adam
-#    optim_conf:
-#        lr: 10.
-#    scheduler: noamlr
-#    scheduler_conf:
-#        model_size: 256
-#        warmup_steps: 25000
-#
-optim: adam
-optim_conf:
-    lr: 0.00395
-scheduler: warmuplr     # pytorch v1.1.0+ required
-scheduler_conf:
-    warmup_steps: 25000
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer2.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer2.yaml
deleted file mode 100644
index 1ce32fff48f..00000000000
--- a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer2.yaml
+++ /dev/null
@@ -1,54 +0,0 @@
-# network architecture
-# encoder related
-encoder: transformer
-encoder_conf:
-    output_size: 256    # dimension of attention
-    attention_heads: 4
-    linear_units: 2048  # the number of units of position-wise feed forward
-    num_blocks: 12      # the number of encoder blocks
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    attention_dropout_rate: 0.0
-    input_layer: conv2d # encoder architecture type
-    normalize_before: true
-
-# decoder related
-decoder: transformer
-decoder_conf:
-    attention_heads: 4
-    linear_units: 2048
-    num_blocks: 6
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    self_attention_dropout_rate: 0.0
-    src_attention_dropout_rate: 0.0
-
-# hybrid CTC/attention
-model_conf:
-    ctc_weight: 0.3
-    lsm_weight: 0.1     # label smoothing option
-    length_normalized_loss: false
-
-# minibatch related
-batch_type: folded
-batch_size: 32
-
-# optimization related
-accum_grad: 2
-grad_clip: 5
-max_epoch: 50
-val_scheduler_criterion:
-    - valid
-    - acc
-best_model_criterion:
--   - valid
-    - acc
-    - max
-keep_nbest_models: 10
-
-optim: adam
-optim_conf:
-   lr: 0.0004
-scheduler: warmuplr
-scheduler_conf:
-   warmup_steps: 25000
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer3.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer3.yaml
deleted file mode 100644
index 523c2bc8e84..00000000000
--- a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer3.yaml
+++ /dev/null
@@ -1,70 +0,0 @@
-# network architecture
-# encoder related
-encoder: transformer
-encoder_conf:
-    output_size: 256    # dimension of attention
-    attention_heads: 4
-    linear_units: 2048  # the number of units of position-wise feed forward
-    num_blocks: 12      # the number of encoder blocks
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    attention_dropout_rate: 0.0
-    input_layer: conv2d # encoder architecture type
-    normalize_before: true
-
-# decoder related
-decoder: transformer
-decoder_conf:
-    attention_heads: 4
-    linear_units: 2048
-    num_blocks: 6
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    self_attention_dropout_rate: 0.0
-    src_attention_dropout_rate: 0.0
-
-# hybrid CTC/attention
-model_conf:
-    ctc_weight: 0.3
-    lsm_weight: 0.1     # label smoothing option
-    length_normalized_loss: false
-
-# minibatch related
-batch_type: numel
-batch_bins: 8500000
-
-# optimization related
-accum_grad: 2
-grad_clip: 5
-max_epoch: 50
-val_scheduler_criterion:
-    - valid
-    - acc
-best_model_criterion:
--   - valid
-    - acc
-    - max
-keep_nbest_models: 10
-
-optim: adam
-optim_conf:
-   lr: 0.0005
-scheduler: warmuplr
-scheduler_conf:
-   warmup_steps: 30000
-
-specaug: specaug
-specaug_conf:
-    apply_time_warp: true
-    time_warp_window: 5
-    time_warp_mode: bicubic
-    apply_freq_mask: true
-    freq_mask_width_range:
-    - 0
-    - 30
-    num_freq_mask: 2
-    apply_time_mask: true
-    time_mask_width_range:
-    - 0
-    - 40
-    num_time_mask: 2
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer_lr0.002.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer_lr0.002.yaml
deleted file mode 100644
index 7097638fcff..00000000000
--- a/egs2/zh_openslr38/asr1/conf/tuning/train_asr_transformer_lr0.002.yaml
+++ /dev/null
@@ -1,66 +0,0 @@
-# network architecture
-# encoder related
-encoder: transformer
-encoder_conf:
-    output_size: 256    # dimension of attention
-    attention_heads: 4
-    linear_units: 2048  # the number of units of position-wise feed forward
-    num_blocks: 12      # the number of encoder blocks
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    attention_dropout_rate: 0.0
-    input_layer: conv2d # encoder architecture type
-    normalize_before: true
-
-# decoder related
-decoder: transformer
-decoder_conf:
-    attention_heads: 4
-    linear_units: 2048
-    num_blocks: 6
-    dropout_rate: 0.1
-    positional_dropout_rate: 0.1
-    self_attention_dropout_rate: 0.0
-    src_attention_dropout_rate: 0.0
-
-# hybrid CTC/attention
-model_conf:
-    ctc_weight: 0.3
-    lsm_weight: 0.1     # label smoothing option
-    length_normalized_loss: false
-
-# minibatch related
-batch_type: folded
-batch_size: 32
-
-# optimization related
-accum_grad: 2
-grad_clip: 5
-patience: 3
-max_epoch: 20
-val_scheduler_criterion:
-    - valid
-    - acc
-best_model_criterion:
--   - valid
-    - acc
-    - max
-keep_nbest_models: 10
-
-# NoamLR is deprecated. Use WarmupLR.
-# The following is equivalent setting for NoamLR:
-#
-#    optim: adam
-#    optim_conf:
-#        lr: 10.
-#    scheduler: noamlr
-#    scheduler_conf:
-#        model_size: 256
-#        warmup_steps: 25000
-#
-optim: adam
-optim_conf:
-    lr: 0.002
-scheduler: warmuplr     # pytorch v1.1.0+ required
-scheduler_conf:
-    warmup_steps: 25000
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn.yaml
deleted file mode 100644
index 26ea1e4f96e..00000000000
--- a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn.yaml
+++ /dev/null
@@ -1,19 +0,0 @@
-# rnnlm related
-lm: seq_rnn
-lm_conf:
-    unit: 650
-    nlayers: 2
-
-# optimization related
-grad_clip: 5.0
-batch_type: folded
-batch_size: 64  # batch size in LM training
-max_epoch: 20   # if the data size is large, we can reduce this
-patience: 3
-optim: sgd
-
-best_model_criterion:
--   - valid
-    - loss
-    - min
-keep_nbest_models: 1
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn2.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn2.yaml
deleted file mode 100644
index b7459681adf..00000000000
--- a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_rnn2.yaml
+++ /dev/null
@@ -1,19 +0,0 @@
-# rnnlm related
-lm: seq_rnn
-lm_conf:
-    unit: 650
-    nlayers: 2
-
-# optimization related
-grad_clip: 5.0
-batch_type: folded
-batch_size: 64  # batch size in LM training
-max_epoch: 40   # if the data size is large, we can reduce this
-patience: 3
-optim: sgd
-
-best_model_criterion:
--   - valid
-    - loss
-    - min
-keep_nbest_models: 1
diff --git a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_transformer.yaml b/egs2/zh_openslr38/asr1/conf/tuning/train_lm_transformer.yaml
deleted file mode 100644
index ace0739a939..00000000000
--- a/egs2/zh_openslr38/asr1/conf/tuning/train_lm_transformer.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-lm: transformer
-lm_conf:
-    pos_enc: null
-    embed_unit: 128
-    att_unit: 512
-    head: 8
-    unit: 2048
-    layer: 16
-    dropout_rate: 0.1
-
-# optimization related
-grad_clip: 5.0
-batch_type: numel
-batch_bins: 2000000
-accum_grad: 1
-max_epoch: 15  # 15epoch is enougth
-
-optim: adam
-optim_conf:
-   lr: 0.001
-scheduler: warmuplr
-scheduler_conf:
-   warmup_steps: 25000
-
-best_model_criterion:
--   - valid
-    - loss
-    - min
-keep_nbest_models: 10  # 10 is good.
diff --git a/egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py b/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py
similarity index 100%
rename from egs2/TEMPLATE/asr1/pyscripts/utils/check_train_test_duplicate.py
rename to egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py
diff --git a/egs2/zh_openslr38/asr1/local/data.sh b/egs2/zh_openslr38/asr1/local/data.sh
index 15c0341c7eb..f528e558403 100755
--- a/egs2/zh_openslr38/asr1/local/data.sh
+++ b/egs2/zh_openslr38/asr1/local/data.sh
@@ -77,7 +77,7 @@ for dir in $train_dir $dev_dir $test_dir; do
   utils/utt2spk_to_spk2utt.pl $dir/utt2spk > $dir/spk2utt
 done
 
-python3 pyscripts/utils/check_train_test_duplicate.py
+python3 local/check_train_test_duplicate.py
 
 # validate formats
 utils/validate_data_dir.sh --no-feats data/train
diff --git a/egs2/zh_openslr38/asr1/local/download_and_untar.sh b/egs2/zh_openslr38/asr1/local/download_and_untar.sh
deleted file mode 120000
index 5e67b127005..00000000000
--- a/egs2/zh_openslr38/asr1/local/download_and_untar.sh
+++ /dev/null
@@ -1 +0,0 @@
-../../../../egs/aishell/asr1/local/download_and_untar.sh
\ No newline at end of file
diff --git a/egs2/zh_openslr38/asr1/run.sh b/egs2/zh_openslr38/asr1/run.sh
index fb1d737316a..8d443a09702 100755
--- a/egs2/zh_openslr38/asr1/run.sh
+++ b/egs2/zh_openslr38/asr1/run.sh
@@ -9,10 +9,10 @@ train_set=train
 valid_set=dev
 test_sets="dev test"
 
-asr_config=conf/train_asr_conformer.yaml
-inference_config=conf/decode_asr_rnn.yaml
+asr_config=conf/train_asr.yaml
+inference_config=conf/decode_asr.yaml
 
-lm_config=conf/train_lm_transformer.yaml
+lm_config=conf/train_lm.yaml
 use_lm=true
 use_wordlm=false
 
diff --git a/egs2/zh_openslr38/asr1/run_sslr.sh b/egs2/zh_openslr38/asr1/run_sslr.sh
index 1e391b9379e..0f924c725c5 100755
--- a/egs2/zh_openslr38/asr1/run_sslr.sh
+++ b/egs2/zh_openslr38/asr1/run_sslr.sh
@@ -9,10 +9,10 @@ train_set=train
 valid_set=dev
 test_sets="dev test"
 
-asr_config=conf/train_asr_conformer_sslr.yaml
-inference_config=conf/decode_asr_rnn.yaml
+asr_config=conf/tuning/train_asr_sslr.yaml
+inference_config=conf/decode_asr.yaml
 
-lm_config=conf/train_lm_transformer.yaml
+lm_config=conf/train_lm.yaml
 use_lm=true
 use_wordlm=false
 
@@ -40,4 +40,5 @@ speed_perturb_factors="0.9 1.0 1.1"
     --lm_train_text "data/${train_set}/text" "$@" \
     --feats_normalize uttmvn \
     --nj 1 \
-    --stage 10
+    --inference_asr_model valid.acc.best.pth \
+    --gpu_inference true
diff --git a/egs2/zh_openslr38/asr1/run_streaming.sh b/egs2/zh_openslr38/asr1/run_streaming.sh
deleted file mode 100755
index 8149ae6e842..00000000000
--- a/egs2/zh_openslr38/asr1/run_streaming.sh
+++ /dev/null
@@ -1,41 +0,0 @@
-#!/usr/bin/env bash
-# Set bash to 'debug' mode, it will exit on :
-# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
-set -e
-set -u
-set -o pipefail
-
-train_set=train
-valid_set=dev
-test_sets="dev test"
-
-asr_config=conf/train_asr_streaming_transformer.yaml
-inference_config=conf/decode_asr_streaming.yaml
-
-lm_config=conf/train_lm.yaml
-use_lm=true
-use_wordlm=false
-
-# speed perturbation related
-# (train_set will be "${train_set}_sp" if speed_perturb_factors is specified)
-speed_perturb_factors="0.9 1.0 1.1"
-
-./asr.sh                                               \
-    --use_streaming true                               \
-    --lang zh                                          \
-    --audio_format wav                                 \
-    --feats_type raw                                   \
-    --token_type char                                  \
-    --use_lm ${use_lm}                                 \
-    --use_word_lm ${use_wordlm}                        \
-    --lm_config "${lm_config}"                         \
-    --asr_config "${asr_config}"                       \
-    --inference_config "${inference_config}"           \
-    --train_set "${train_set}"                         \
-    --valid_set "${valid_set}"                         \
-    --test_sets "${test_sets}"                         \
-    --speed_perturb_factors "${speed_perturb_factors}" \
-    --asr_speech_fold_length 512 \
-    --asr_text_fold_length 150 \
-    --lm_fold_length 150 \
-    --lm_train_text "data/${train_set}/text" "$@"

From 2953b4e76ceddaa5d95bf37f06fcd9f229bff3b2 Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-18-124.us-east-2.compute.internal>
Date: Tue, 22 Mar 2022 00:21:08 +0000
Subject: [PATCH 12/22] reformat python files with black

---
 .../asr1/local/check_train_test_duplicate.py  |  23 ++--
 egs2/zh_openslr38/asr1/local/data_split.py    | 103 ++++++++++--------
 2 files changed, 69 insertions(+), 57 deletions(-)

diff --git a/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py b/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py
index cb1406bee54..4a2720e777f 100644
--- a/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py
+++ b/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py
@@ -1,11 +1,12 @@
-from collections import Counter 
+from collections import Counter
 
 train_file = "data/train/text"
 
 train_lines = []
 with open(train_file) as f:
     for line in f:
-        if not line: continue
+        if not line:
+            continue
         train_lines.append(line.split()[1])
 train_lines = set(train_lines)
 
@@ -16,7 +17,8 @@
     test_uttids = []
     with open(test_file) as f:
         for line in f:
-            if not line: continue
+            if not line:
+                continue
             test_uttids.append(line.split()[0])
             test_lines.append(line.split()[1])
 
@@ -31,18 +33,19 @@
 
     # if input("continue? [y/n]") == 'y':
     # remove all instances of duplicate uttids in these files: spk2utt, text, utt2spk, wav.scp
-    with open(f"data/{test_name}/spk2utt", 'r') as f:
+    with open(f"data/{test_name}/spk2utt", "r") as f:
         # replace all uttid with empty string
         text = f.read()
         for uttid in duplicate_uttids:
-            text = text.replace(" "+uttid, "")
+            text = text.replace(" " + uttid, "")
         for line in text.split("\n"):
-            if not line: continue
+            if not line:
+                continue
             if len(line.strip().split(" ")) < 2:
                 print(f"removing {line} from spk2utt")
-                text = text.replace(line+'\n', "")
-    with open(f"data/{test_name}/spk2utt", 'w') as f:
-        f.write(text)   
+                text = text.replace(line + "\n", "")
+    with open(f"data/{test_name}/spk2utt", "w") as f:
+        f.write(text)
 
     for name in ("text", "utt2spk", "wav.scp"):
         with open(f"data/{test_name}/{name}", "r") as f:
@@ -55,4 +58,4 @@
             f.write("\n".join(out_lines))
             f.write("\n")
     # else:
-    #     print("ok.")
\ No newline at end of file
+    #     print("ok.")
diff --git a/egs2/zh_openslr38/asr1/local/data_split.py b/egs2/zh_openslr38/asr1/local/data_split.py
index 60b351f6965..df952d304cd 100644
--- a/egs2/zh_openslr38/asr1/local/data_split.py
+++ b/egs2/zh_openslr38/asr1/local/data_split.py
@@ -1,4 +1,3 @@
-
 """
 Split data to train, dev, test
 """
@@ -10,80 +9,90 @@
 train_size = 0.9
 random.seed(1)
 
-data_dir = sys.argv[1] #ST-CMDS-20170001_1-OS
+data_dir = sys.argv[1]  # ST-CMDS-20170001_1-OS
 
 # create speaker id dictionary
 d = defaultdict(list)
 for fn in os.listdir(data_dir):
-	if not fn.endswith('.wav'): continue
-	# 20170001P00001A0001.wav
-	prefix, s = fn.split('P')
-	try:
-		speaker, s = s.split('A')
-		letter = 'A'
-	except ValueError:
-		speaker, s = s.split('I')
-		letter = 'I'
-	utt, _ = s.split('.')
-	d[speaker+letter].append(utt)
+    if not fn.endswith(".wav"):
+        continue
+    # 20170001P00001A0001.wav
+    prefix, s = fn.split("P")
+    try:
+        speaker, s = s.split("A")
+        letter = "A"
+    except ValueError:
+        speaker, s = s.split("I")
+        letter = "I"
+    utt, _ = s.split(".")
+    d[speaker + letter].append(utt)
 
 speaker_ids = list(d.keys())
 random.shuffle(speaker_ids)
 
 num_speakers = len(speaker_ids)
-assert num_speakers == 855, "Number of speakers should be 855 in Free ST Chinese Mandarin Corpus."
+assert (
+    num_speakers == 855
+), "Number of speakers should be 855 in Free ST Chinese Mandarin Corpus."
 
-num_train = int(train_size*num_speakers)
-num_test = int((num_speakers - num_train)/2)
+num_train = int(train_size * num_speakers)
+num_test = int((num_speakers - num_train) / 2)
 
 train_speakers = speaker_ids[:num_train]
 dev_speakers = speaker_ids[num_train:-num_test]
 test_speakers = speaker_ids[-num_test:]
 
-print(f"# train: {num_train}, # dev:{num_speakers-num_train-num_test}, # test:{num_test}")
+print(
+    f"# train: {num_train}, # dev:{num_speakers-num_train-num_test}, # test:{num_test}"
+)
+
 
 def get_transcription(spk_id, utt_id):
-	text_fn = get_text_filename(spk_id, utt_id)
-	with open(text_fn) as f:
-		lines = f.readlines()
-	assert len(lines) == 1, f"More than one line in transription file:{text_fn}"
-	return lines[0]
+    text_fn = get_text_filename(spk_id, utt_id)
+    with open(text_fn) as f:
+        lines = f.readlines()
+    assert len(lines) == 1, f"More than one line in transription file:{text_fn}"
+    return lines[0]
+
 
 def get_text_filename(spk_id, utt_id):
-	return f"{data_dir}/20170001P{spk_id}{utt_id}.txt"
+    return f"{data_dir}/20170001P{spk_id}{utt_id}.txt"
+
 
 def get_wav_filename(spk_id, utt_id):
-	return f"{data_dir}/20170001P{spk_id}{utt_id}.wav"
+    return f"{data_dir}/20170001P{spk_id}{utt_id}.wav"
+
 
 def create_files(speakers, directory):
-	text_lines, scp_lines, utt2spk_lines = [], [], []
-	for spk_id in speakers:
-		for utt_id in d[spk_id]:
-			# add spk_id in front to make utt_id unique
-			unique_utt_id = spk_id+utt_id
+    text_lines, scp_lines, utt2spk_lines = [], [], []
+    for spk_id in speakers:
+        for utt_id in d[spk_id]:
+            # add spk_id in front to make utt_id unique
+            unique_utt_id = spk_id + utt_id
+
+            transcription = get_transcription(spk_id, utt_id)
+            text_lines.append(f"{unique_utt_id} {transcription}\n")
 
-			transcription = get_transcription(spk_id, utt_id)
-			text_lines.append(f"{unique_utt_id} {transcription}\n")
+            wav_file_path = get_wav_filename(spk_id, utt_id)
+            scp_lines.append(f"{unique_utt_id} {wav_file_path}\n")
 
-			wav_file_path = get_wav_filename(spk_id, utt_id)
-			scp_lines.append(f"{unique_utt_id} {wav_file_path}\n")
+            utt2spk_lines.append(f"{unique_utt_id} {spk_id}\n")
 
-			utt2spk_lines.append(f"{unique_utt_id} {spk_id}\n")
+    # sort
+    text_lines.sort()
+    scp_lines.sort()
+    utt2spk_lines.sort()
 
-	# sort
-	text_lines.sort()
-	scp_lines.sort()
-	utt2spk_lines.sort()
+    # write to file
+    with open(f"{directory}/text", "w+") as text_file:
+        text_file.writelines(text_lines)
 
-	# write to file
-	with open(f"{directory}/text", "w+") as text_file:
-		text_file.writelines(text_lines)
+    with open(f"{directory}/wav.scp", "w+") as scp_file:
+        scp_file.writelines(scp_lines)
+
+    with open(f"{directory}/utt2spk", "w+") as utt2spk_file:
+        utt2spk_file.writelines(utt2spk_lines)
 
-	with open(f"{directory}/wav.scp", "w+") as scp_file:
-		scp_file.writelines(scp_lines)
-	
-	with open(f"{directory}/utt2spk", "w+") as utt2spk_file:
-		utt2spk_file.writelines(utt2spk_lines)
 
 print("Creating files for train...", end="")
 create_files(train_speakers, "data/train")
@@ -95,4 +104,4 @@ def create_files(speakers, directory):
 
 print("Creating files for test...", end="")
 create_files(test_speakers, "data/test")
-print("Done.")
\ No newline at end of file
+print("Done.")

From ed904c9aa3d2bafd69a5970238d9a363c3c811c5 Mon Sep 17 00:00:00 2001
From: Chen Cui <cxcui@andrew.cmu.edu>
Date: Tue, 22 Mar 2022 00:57:21 +0000
Subject: [PATCH 13/22] Add corpus to egs2

---
 egs2/README.md | 1 +
 1 file changed, 1 insertion(+)

diff --git a/egs2/README.md b/egs2/README.md
index d03884f2ceb..e7669c6fc69 100755
--- a/egs2/README.md
+++ b/egs2/README.md
@@ -93,3 +93,4 @@ See: https://espnet.github.io/espnet/espnet2_tutorial.html#recipes-using-espnet2
 | yesno                   | The "yesno" corpus                                                                      | ASR                     | HEB                  | http://www.openslr.org/1                                                                                     |              |
 | yoloxochitl_mixtec      | Yoloxochitl-Mixtec corpus (endangered language in central Mexico)                       | ASR                     | XTY                  | http://www.openslr.org/89                                                                                    |              |
 | zeroth_korean           | Zeroth-Korean                                                                           | ASR                     | KOR                  | http://www.openslr.org/40                                                                                    |              |
+| zh_openslr38            | ST-CMDS-20170001_1, Free ST Chinese Mandarin Corpus                                     | ASR                     | CMN                  | http://www.openslr.org/38                                                                                    |              |
\ No newline at end of file

From 148018e3ac257f44c16d954ca9839ca380ab95d7 Mon Sep 17 00:00:00 2001
From: Chen Cui <cxcui@andrew.cmu.edu>
Date: Tue, 22 Mar 2022 01:04:47 +0000
Subject: [PATCH 14/22] Add huggingface pretrained model link

---
 egs2/zh_openslr38/asr1/README.md | 1 +
 1 file changed, 1 insertion(+)

diff --git a/egs2/zh_openslr38/asr1/README.md b/egs2/zh_openslr38/asr1/README.md
index 40fa2bc49d6..e106febf202 100644
--- a/egs2/zh_openslr38/asr1/README.md
+++ b/egs2/zh_openslr38/asr1/README.md
@@ -12,6 +12,7 @@ Link: https://openslr.org/38
 - python version: `3.9.10 | packaged by conda-forge | (main, Feb  1 2022, 21:24:11)  [GCC 9.4.0]`
 - espnet version: `espnet 0.10.7a1`
 - pytorch version: `pytorch 1.10.1`
+- pretrained model: https://huggingface.co/espnet/zh_openslr38/blob/main/exp/asr_train_asr_conformer_raw_zh_char_sp/valid.acc.ave_10best.pth
 
 ### CER
 

From 7e5ea0967547925aff7af291636304cdc30e67ab Mon Sep 17 00:00:00 2001
From: Chen Cui <cxcui@andrew.cmu.edu>
Date: Wed, 23 Mar 2022 14:31:15 +0000
Subject: [PATCH 15/22] remove unused variables

---
 egs2/zh_openslr38/asr1/local/data.sh | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/egs2/zh_openslr38/asr1/local/data.sh b/egs2/zh_openslr38/asr1/local/data.sh
index f528e558403..a8ae818556d 100755
--- a/egs2/zh_openslr38/asr1/local/data.sh
+++ b/egs2/zh_openslr38/asr1/local/data.sh
@@ -59,9 +59,6 @@ ST_CMDS=$(cd ${ST_CMDS}; pwd)
 echo local/data_download.sh ${download_opt} "${ST_CMDS}" "${data_url}" ST-CMDS-20170001_1-OS.tar.gz
 local/data_download.sh ${download_opt} "${ST_CMDS}" "${data_url}" ST-CMDS-20170001_1-OS.tar.gz
 
-ST_CMDS_audio_dir=${ST_CMDS}/data_ST_CMDS/wav
-ST_CMDS_text=${ST_CMDS}/data_ST_CMDS/transcript/ST_CMDS_transcript_v0.8.txt
-
 log "Data Preparation"
 train_dir=data/train
 dev_dir=data/dev

From 8e66d8d2dd5017aa7516a270f561afa392d85733 Mon Sep 17 00:00:00 2001
From: cuichenx <bob.cui@mail.utoronto.ca>
Date: Thu, 24 Mar 2022 10:12:53 -0400
Subject: [PATCH 16/22] move alternate run script to local folder and update
 readme

---
 egs2/zh_openslr38/asr1/README.md               | 17 +++++++++++++++++
 egs2/zh_openslr38/asr1/{ => local}/run_sslr.sh |  0
 2 files changed, 17 insertions(+)
 rename egs2/zh_openslr38/asr1/{ => local}/run_sslr.sh (100%)

diff --git a/egs2/zh_openslr38/asr1/README.md b/egs2/zh_openslr38/asr1/README.md
index e106febf202..e646b431596 100644
--- a/egs2/zh_openslr38/asr1/README.md
+++ b/egs2/zh_openslr38/asr1/README.md
@@ -14,9 +14,26 @@ Link: https://openslr.org/38
 - pytorch version: `pytorch 1.10.1`
 - pretrained model: https://huggingface.co/espnet/zh_openslr38/blob/main/exp/asr_train_asr_conformer_raw_zh_char_sp/valid.acc.ave_10best.pth
 
+## Spectrum Features
+Code to reproduce:
+```./run.sh```
+
 ### CER
 
 |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
 |---|---|---|---|---|---|---|---|---|
 |decode_asr_rnn_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/dev|4322|46490|91.0|8.4|0.5|0.2|9.2|51.5|
 |decode_asr_rnn_lm_lm_train_lm_transformer_zh_char_valid.loss.ave_asr_model_valid.acc.ave/test|4167|45803|91.1|8.5|0.5|0.2|9.1|52.2|
+
+## HuBERT Self-Supervised Learning (SSLR)
+We provide the script to train with SSLR features via HuBERT. Due to the much longer training time with HuBERT, we only train for 24 epochs. The model does not show a lower CER over spectrum features, but training for longer may lead to improved results.
+
+Code to reproduce:
+```./local/run_sslr.sh```
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_lm_lm_train_lm_zh_char_valid.loss.ave_asr_model_valid.acc.best/dev|4322|46490|90.8|8.6|0.6|0.2|9.4|51.9|
+|decode_asr_lm_lm_train_lm_zh_char_valid.loss.ave_asr_model_valid.acc.best/test|4167|45803|90.8|8.7|0.5|0.2|9.4|54.1|
diff --git a/egs2/zh_openslr38/asr1/run_sslr.sh b/egs2/zh_openslr38/asr1/local/run_sslr.sh
similarity index 100%
rename from egs2/zh_openslr38/asr1/run_sslr.sh
rename to egs2/zh_openslr38/asr1/local/run_sslr.sh

From 0799274bcc68101cdef39920ab971df4bf97fec0 Mon Sep 17 00:00:00 2001
From: Chaitanya Narisetty <cnariset@andrew.cmu.edu>
Date: Thu, 24 Mar 2022 12:25:15 -0400
Subject: [PATCH 17/22] Fetch upstream (#8)


From 8faedd4489e3af81d375b4ad041e3a00d86df705 Mon Sep 17 00:00:00 2001
From: Chaitanya Narisetty <cnariset@andrew.cmu.edu>
Date: Thu, 24 Mar 2022 15:09:56 -0400
Subject: [PATCH 18/22] black check local data prep

---
 egs2/ms_indic_18/asr1/local/prepare_data.py | 36 ++++++++++-----------
 1 file changed, 17 insertions(+), 19 deletions(-)

diff --git a/egs2/ms_indic_18/asr1/local/prepare_data.py b/egs2/ms_indic_18/asr1/local/prepare_data.py
index 488cd920ff6..464a1f43b11 100755
--- a/egs2/ms_indic_18/asr1/local/prepare_data.py
+++ b/egs2/ms_indic_18/asr1/local/prepare_data.py
@@ -31,20 +31,20 @@
 
 
 # prepare data for training and validation splits
-with open(traindir+'transcription.txt') as f:
+with open(traindir + "transcription.txt") as f:
     train_lines = [line.rstrip() for line in f.readlines()]
     train_id2text = {}
     train_id2filepath = {}
     for line in train_lines:
         wav_id = line.split()[0]
         filepath = f"{traindir}/Audios/{wav_id}.wav"
-        train_id2text[wav_id] = ' '.join(line.split()[1:])
+        train_id2text[wav_id] = " ".join(line.split()[1:])
         train_id2filepath[wav_id] = filepath
 
 wav_ids = list(train_id2text.keys())
 random.shuffle(wav_ids)
 valid_id2text = {}
-valid_totaldur = 2*60*60 # (in seconds) 2 hours taken for validation split
+valid_totaldur = 2 * 60 * 60  # (in seconds) 2 hours taken for validation split
 for wav_id in wav_ids:
     dur = librosa.get_duration(filename=train_id2filepath[wav_id])
     valid_id2text[wav_id] = train_id2text.pop(wav_id)
@@ -53,55 +53,53 @@
         break
 
 
-with open(train_datadir+'text', 'w') as f:
+with open(train_datadir + "text", "w") as f:
     for wav_id in sorted(train_id2text):
         f.write(f"{lang}_{wav_id} {train_id2text[wav_id]}\n")
-with open(train_datadir+'wav.scp', 'w') as f:
+with open(train_datadir + "wav.scp", "w") as f:
     for wav_id in sorted(train_id2text):
         f.write(f"{lang}_{wav_id} {train_id2filepath[wav_id]}\n")
-with open(train_datadir+'spk2utt', 'w') as f:
+with open(train_datadir + "spk2utt", "w") as f:
     for wav_id in sorted(train_id2text):
         f.write(f"spk_{lang}_{wav_id} {lang}_{wav_id}\n")
-with open(train_datadir+'utt2spk', 'w') as f:
+with open(train_datadir + "utt2spk", "w") as f:
     for wav_id in sorted(train_id2text):
         f.write(f"{lang}_{wav_id} spk_{lang}_{wav_id}\n")
 
-with open(valid_datadir+'text', 'w') as f:
+with open(valid_datadir + "text", "w") as f:
     for wav_id in sorted(valid_id2text):
         f.write(f"{lang}_{wav_id} {valid_id2text[wav_id]}\n")
-with open(valid_datadir+'wav.scp', 'w') as f:
+with open(valid_datadir + "wav.scp", "w") as f:
     for wav_id in sorted(valid_id2text):
         f.write(f"{lang}_{wav_id} {train_id2filepath[wav_id]}\n")
-with open(valid_datadir+'spk2utt', 'w') as f:
+with open(valid_datadir + "spk2utt", "w") as f:
     for wav_id in sorted(valid_id2text):
         f.write(f"spk_{lang}_{wav_id} {lang}_{wav_id}\n")
-with open(valid_datadir+'utt2spk', 'w') as f:
+with open(valid_datadir + "utt2spk", "w") as f:
     for wav_id in sorted(valid_id2text):
         f.write(f"{lang}_{wav_id} spk_{lang}_{wav_id}\n")
 
 
 # prepare test data
-with open(testdir+'transcription.txt') as f:
+with open(testdir + "transcription.txt") as f:
     test_lines = [line.rstrip() for line in f.readlines()]
     test_id2text = {}
     test_id2filepath = {}
     for line in test_lines:
         wav_id = line.split()[0]
         filepath = f"{testdir}/Audios/{wav_id}.wav"
-        test_id2text[wav_id] = ' '.join(line.split()[1:])
+        test_id2text[wav_id] = " ".join(line.split()[1:])
         test_id2filepath[wav_id] = filepath
 
-with open(test_datadir+'text', 'w') as f:
+with open(test_datadir + "text", "w") as f:
     for wav_id in sorted(test_id2text):
         f.write(f"{lang}_{wav_id} {test_id2text[wav_id]}\n")
-with open(test_datadir+'wav.scp', 'w') as f:
+with open(test_datadir + "wav.scp", "w") as f:
     for wav_id in sorted(test_id2text):
         f.write(f"{lang}_{wav_id} {test_id2filepath[wav_id]}\n")
-with open(test_datadir+'spk2utt', 'w') as f:
+with open(test_datadir + "spk2utt", "w") as f:
     for wav_id in sorted(test_id2text):
         f.write(f"spk_{lang}_{wav_id} {lang}_{wav_id}\n")
-with open(test_datadir+'utt2spk', 'w') as f:
+with open(test_datadir + "utt2spk", "w") as f:
     for wav_id in sorted(test_id2text):
         f.write(f"{lang}_{wav_id} spk_{lang}_{wav_id}\n")
-
-

From d24d64c4c7210baeb7dccffd7d774e30374c7d38 Mon Sep 17 00:00:00 2001
From: Chaitanya Narisetty <cnariset@andrew.cmu.edu>
Date: Thu, 24 Mar 2022 15:16:12 -0400
Subject: [PATCH 19/22] Update README.md

Fixed name of recipe: `ms_indic_18`
---
 egs2/README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/egs2/README.md b/egs2/README.md
index f11f2b51e45..6db3e3e0930 100755
--- a/egs2/README.md
+++ b/egs2/README.md
@@ -57,7 +57,7 @@ See: https://espnet.github.io/espnet/espnet2_tutorial.html#recipes-using-espnet2
 | mini_an4                | Mini version of CMU AN4 database for the integration test                               | ASR/TTS/SE              | ENG                  | http://www.speech.cs.cmu.edu/databases/an4/                                                                  |              |
 | mini_librispeech        | Mini version of Librispeech corpus                                                      | DIAR                    | ENG                  | https://openslr.org/31/                                                                                      |              |
 | mls                     | MLS (A large multilingual corpus derived from LibriVox audiobooks)                      | ASR                     | 8 languages          | http://www.openslr.org/94/                                                                                   |              |
-| ms_indic_is18           | Microsoft Speech Corpus (Indian languages)                                              | ASR                     | 3 langs: TEL TAM GUJ | https://msropendata.com/datasets/7230b4b1-912d-400e-be58-f84e0512985e                                        |              |
+| ms_indic_18             | Microsoft Speech Corpus (Indian languages)                                              | ASR                     | 3 langs: TEL TAM GUJ | https://msropendata.com/datasets/7230b4b1-912d-400e-be58-f84e0512985e                                        |              |
 | nsc                     | National Speech Corpus                                                                  | ASR                     | ENG-SG               | https://www.imda.gov.sg/programme-listing/digital-services-lab/national-speech-corpus                        |              |
 | open_li52               | Corpus combination with 52 languages(Commonvocie + voxforge)                            | Multilingual ASR        | 52 languages         |                                                                                                             |              |
 | polyphone_swiss_french  | Swiss French Polyphone corpus                                                           | ASR                     | FRA                  | http://catalog.elra.info/en-us/repository/browse/ELRA-S0030_02                                               |              |

From b71c374f5c06c9227e46994ea7d516f4058aa6cd Mon Sep 17 00:00:00 2001
From: Chaitanya Narisetty <cnariset@andrew.cmu.edu>
Date: Thu, 24 Mar 2022 16:44:08 -0400
Subject: [PATCH 20/22] Updated links to config links

---
 egs2/ms_indic_18/asr1/conf/decode_asr.yaml | 2 +-
 egs2/ms_indic_18/asr1/conf/train_asr.yaml  | 2 +-
 egs2/ms_indic_18/asr1/conf/train_lm.yaml   | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/egs2/ms_indic_18/asr1/conf/decode_asr.yaml b/egs2/ms_indic_18/asr1/conf/decode_asr.yaml
index 058d7b2c88f..78955c67707 120000
--- a/egs2/ms_indic_18/asr1/conf/decode_asr.yaml
+++ b/egs2/ms_indic_18/asr1/conf/decode_asr.yaml
@@ -1 +1 @@
-conf/tuning/decode_asr_transformer.yaml
\ No newline at end of file
+tuning/decode_asr_transformer.yaml
\ No newline at end of file
diff --git a/egs2/ms_indic_18/asr1/conf/train_asr.yaml b/egs2/ms_indic_18/asr1/conf/train_asr.yaml
index 3ca87b197da..00ea2f24791 120000
--- a/egs2/ms_indic_18/asr1/conf/train_asr.yaml
+++ b/egs2/ms_indic_18/asr1/conf/train_asr.yaml
@@ -1 +1 @@
-conf/tuning/train_asr_transformer.yaml
\ No newline at end of file
+tuning/train_asr_transformer.yaml
\ No newline at end of file
diff --git a/egs2/ms_indic_18/asr1/conf/train_lm.yaml b/egs2/ms_indic_18/asr1/conf/train_lm.yaml
index 2be221361aa..dd50b722db8 120000
--- a/egs2/ms_indic_18/asr1/conf/train_lm.yaml
+++ b/egs2/ms_indic_18/asr1/conf/train_lm.yaml
@@ -1 +1 @@
-conf/tuning/train_lm_rnn.yaml
\ No newline at end of file
+tuning/train_lm_rnn.yaml
\ No newline at end of file

From fb084d38aeae301465cb93528913f3c2ce420e4a Mon Sep 17 00:00:00 2001
From: Ubuntu <ubuntu@ip-172-31-16-217.us-east-2.compute.internal>
Date: Mon, 21 Mar 2022 00:40:52 +0000
Subject: [PATCH 21/22] Added Marathi OpenSLR 64 recipe

---
 egs2/README.md                                |  3 +-
 egs2/TEMPLATE/asr1/db.sh                      |  1 +
 egs2/mr_openslr64/asr1/README.md              | 36 +++++++
 egs2/mr_openslr64/asr1/asr.sh                 |  1 +
 egs2/mr_openslr64/asr1/cmd.sh                 |  1 +
 egs2/mr_openslr64/asr1/conf/decode_asr.yaml   |  1 +
 egs2/mr_openslr64/asr1/conf/fbank.conf        |  2 +
 egs2/mr_openslr64/asr1/conf/pbs.conf          | 11 +++
 egs2/mr_openslr64/asr1/conf/pitch.conf        |  1 +
 egs2/mr_openslr64/asr1/conf/queue.conf        | 12 +++
 egs2/mr_openslr64/asr1/conf/slurm.conf        | 14 +++
 egs2/mr_openslr64/asr1/conf/train_asr.yaml    |  1 +
 egs2/mr_openslr64/asr1/conf/train_lm.yaml     | 14 +++
 .../asr1/conf/tuning/decode_rnn.yaml          |  6 ++
 .../asr1/conf/tuning/decode_transformer.yaml  |  7 ++
 .../asr1/conf/tuning/train_asr_conformer.yaml | 67 +++++++++++++
 .../conf/tuning/train_asr_conformer_xlsr.yaml | 88 +++++++++++++++++
 .../asr1/conf/tuning/train_asr_rnn.yaml       | 35 +++++++
 .../conf/tuning/train_asr_transformer.yaml    | 68 +++++++++++++
 egs2/mr_openslr64/asr1/db.sh                  |  1 +
 egs2/mr_openslr64/asr1/local/data.sh          | 63 ++++++++++++
 egs2/mr_openslr64/asr1/local/data_prep.py     | 99 +++++++++++++++++++
 egs2/mr_openslr64/asr1/local/path.sh          |  0
 egs2/mr_openslr64/asr1/path.sh                |  1 +
 egs2/mr_openslr64/asr1/pyscripts              |  1 +
 egs2/mr_openslr64/asr1/run.sh                 | 41 ++++++++
 egs2/mr_openslr64/asr1/scripts                |  1 +
 egs2/mr_openslr64/asr1/steps                  |  1 +
 egs2/mr_openslr64/asr1/utils                  |  1 +
 29 files changed, 577 insertions(+), 1 deletion(-)
 create mode 100644 egs2/mr_openslr64/asr1/README.md
 create mode 120000 egs2/mr_openslr64/asr1/asr.sh
 create mode 120000 egs2/mr_openslr64/asr1/cmd.sh
 create mode 120000 egs2/mr_openslr64/asr1/conf/decode_asr.yaml
 create mode 100644 egs2/mr_openslr64/asr1/conf/fbank.conf
 create mode 100644 egs2/mr_openslr64/asr1/conf/pbs.conf
 create mode 100644 egs2/mr_openslr64/asr1/conf/pitch.conf
 create mode 100644 egs2/mr_openslr64/asr1/conf/queue.conf
 create mode 100644 egs2/mr_openslr64/asr1/conf/slurm.conf
 create mode 120000 egs2/mr_openslr64/asr1/conf/train_asr.yaml
 create mode 100644 egs2/mr_openslr64/asr1/conf/train_lm.yaml
 create mode 100644 egs2/mr_openslr64/asr1/conf/tuning/decode_rnn.yaml
 create mode 100644 egs2/mr_openslr64/asr1/conf/tuning/decode_transformer.yaml
 create mode 100644 egs2/mr_openslr64/asr1/conf/tuning/train_asr_conformer.yaml
 create mode 100644 egs2/mr_openslr64/asr1/conf/tuning/train_asr_conformer_xlsr.yaml
 create mode 100644 egs2/mr_openslr64/asr1/conf/tuning/train_asr_rnn.yaml
 create mode 100644 egs2/mr_openslr64/asr1/conf/tuning/train_asr_transformer.yaml
 create mode 120000 egs2/mr_openslr64/asr1/db.sh
 create mode 100755 egs2/mr_openslr64/asr1/local/data.sh
 create mode 100644 egs2/mr_openslr64/asr1/local/data_prep.py
 create mode 100644 egs2/mr_openslr64/asr1/local/path.sh
 create mode 120000 egs2/mr_openslr64/asr1/path.sh
 create mode 120000 egs2/mr_openslr64/asr1/pyscripts
 create mode 100755 egs2/mr_openslr64/asr1/run.sh
 create mode 120000 egs2/mr_openslr64/asr1/scripts
 create mode 120000 egs2/mr_openslr64/asr1/steps
 create mode 120000 egs2/mr_openslr64/asr1/utils

diff --git a/egs2/README.md b/egs2/README.md
index 6db3e3e0930..6c4dd142863 100755
--- a/egs2/README.md
+++ b/egs2/README.md
@@ -57,7 +57,8 @@ See: https://espnet.github.io/espnet/espnet2_tutorial.html#recipes-using-espnet2
 | mini_an4                | Mini version of CMU AN4 database for the integration test                               | ASR/TTS/SE              | ENG                  | http://www.speech.cs.cmu.edu/databases/an4/                                                                  |              |
 | mini_librispeech        | Mini version of Librispeech corpus                                                      | DIAR                    | ENG                  | https://openslr.org/31/                                                                                      |              |
 | mls                     | MLS (A large multilingual corpus derived from LibriVox audiobooks)                      | ASR                     | 8 languages          | http://www.openslr.org/94/                                                                                   |              |
-| ms_indic_18             | Microsoft Speech Corpus (Indian languages)                                              | ASR                     | 3 langs: TEL TAM GUJ | https://msropendata.com/datasets/7230b4b1-912d-400e-be58-f84e0512985e                                        |              |
+| mr_openslr64            | OpenSLR Marathi Corpus                                                                  | ASR                     | MAR                  | http://www.openslr.org/64/                                                                                   |              |
+| ms_indic_is18           | Microsoft Speech Corpus (Indian languages)                                              | ASR                     | 3 langs: TEL TAM GUJ | https://msropendata.com/datasets/7230b4b1-912d-400e-be58-f84e0512985e                                        |              |
 | nsc                     | National Speech Corpus                                                                  | ASR                     | ENG-SG               | https://www.imda.gov.sg/programme-listing/digital-services-lab/national-speech-corpus                        |              |
 | open_li52               | Corpus combination with 52 languages(Commonvocie + voxforge)                            | Multilingual ASR        | 52 languages         |                                                                                                             |              |
 | polyphone_swiss_french  | Swiss French Polyphone corpus                                                           | ASR                     | FRA                  | http://catalog.elra.info/en-us/repository/browse/ELRA-S0030_02                                               |              |
diff --git a/egs2/TEMPLATE/asr1/db.sh b/egs2/TEMPLATE/asr1/db.sh
index 6d9b35313b2..84091d63f8f 100755
--- a/egs2/TEMPLATE/asr1/db.sh
+++ b/egs2/TEMPLATE/asr1/db.sh
@@ -129,6 +129,7 @@ SEAME=
 BENGALI=downloads
 IWSLT14=
 MS_INDIC_IS18=
+MARATHI=downloads
 
 # For only CMU TIR environment
 if [[ "$(hostname)" == tir* ]]; then
diff --git a/egs2/mr_openslr64/asr1/README.md b/egs2/mr_openslr64/asr1/README.md
new file mode 100644
index 00000000000..0e6848f9c27
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/README.md
@@ -0,0 +1,36 @@
+# RESULTS
+## Environments
+- date: `Mon Mar 21 16:06:03 UTC 2022`
+- python version: `3.9.7 (default, Sep 16 2021, 13:09:58)  [GCC 7.5.0]`
+- espnet version: `espnet 0.10.7a1`
+- pytorch version: `pytorch 1.11.0+cu102`
+- Git hash: `91325a1e58ca0b13494b94bf79b186b095fe0b58`
+  - Commit date: `Mon Mar 21 00:40:52 2022 +0000`
+
+## asr_train_asr_conformer_xlsr_raw_bpe150_sp
+
+This recipe is for the Marathi language and is trained on the [OpenSLR Marathi](https://www.openslr.org/64/) multi-speaker speech data set.
+
+The following results are obtained by using an XLSR frontend.
+
+Train ASR Config: [conf/tuning/train_asr_conformer_xlsr.yaml](conf/tuning/train_asr_conformer_xlsr.yaml)
+
+Trained Model: [espnet/marathi_openslr64](https://huggingface.co/espnet/marathi_openslr64)
+
+### WER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_batch_size1_asr_model_valid.acc.ave/marathi_test|299|3625|72.9|22.5|4.7|1.7|28.9|88.6|
+
+### CER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_batch_size1_asr_model_valid.acc.ave/marathi_test|299|20557|91.4|3.1|5.5|1.9|10.5|88.6|
+
+### TER
+
+|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
+|---|---|---|---|---|---|---|---|---|
+|decode_asr_batch_size1_asr_model_valid.acc.ave/marathi_test|299|13562|86.5|6.3|7.1|1.4|14.9|88.6|
diff --git a/egs2/mr_openslr64/asr1/asr.sh b/egs2/mr_openslr64/asr1/asr.sh
new file mode 120000
index 00000000000..60b05122cfd
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/asr.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/asr.sh
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/cmd.sh b/egs2/mr_openslr64/asr1/cmd.sh
new file mode 120000
index 00000000000..f77e339f822
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/cmd.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/cmd.sh
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/conf/decode_asr.yaml b/egs2/mr_openslr64/asr1/conf/decode_asr.yaml
new file mode 120000
index 00000000000..f3f59d5ac2b
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/decode_asr.yaml
@@ -0,0 +1 @@
+tuning/decode_transformer.yaml
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/conf/fbank.conf b/egs2/mr_openslr64/asr1/conf/fbank.conf
new file mode 100644
index 00000000000..82ac7bd0dbc
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/fbank.conf
@@ -0,0 +1,2 @@
+--sample-frequency=16000 
+--num-mel-bins=80
diff --git a/egs2/mr_openslr64/asr1/conf/pbs.conf b/egs2/mr_openslr64/asr1/conf/pbs.conf
new file mode 100644
index 00000000000..119509938ce
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/pbs.conf
@@ -0,0 +1,11 @@
+# Default configuration
+command qsub -V -v PATH -S /bin/bash
+option name=* -N $0
+option mem=* -l mem=$0
+option mem=0          # Do not add anything to qsub_opts
+option num_threads=* -l ncpus=$0
+option num_threads=1  # Do not add anything to qsub_opts
+option num_nodes=* -l nodes=$0:ppn=1
+default gpu=0
+option gpu=0
+option gpu=* -l ngpus=$0
diff --git a/egs2/mr_openslr64/asr1/conf/pitch.conf b/egs2/mr_openslr64/asr1/conf/pitch.conf
new file mode 100644
index 00000000000..e959a19d5b8
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/pitch.conf
@@ -0,0 +1 @@
+--sample-frequency=16000
diff --git a/egs2/mr_openslr64/asr1/conf/queue.conf b/egs2/mr_openslr64/asr1/conf/queue.conf
new file mode 100644
index 00000000000..500582fab31
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/queue.conf
@@ -0,0 +1,12 @@
+# Default configuration
+command qsub -v PATH -cwd -S /bin/bash -j y -l arch=*64*
+option name=* -N $0
+option mem=* -l mem_free=$0,ram_free=$0
+option mem=0          # Do not add anything to qsub_opts
+option num_threads=* -pe smp $0
+option num_threads=1  # Do not add anything to qsub_opts
+option max_jobs_run=* -tc $0
+option num_nodes=* -pe mpi $0  # You must set this PE as allocation_rule=1
+default gpu=0
+option gpu=0
+option gpu=* -l gpu=$0 -q g.q
diff --git a/egs2/mr_openslr64/asr1/conf/slurm.conf b/egs2/mr_openslr64/asr1/conf/slurm.conf
new file mode 100644
index 00000000000..3b229673638
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/slurm.conf
@@ -0,0 +1,14 @@
+# Default configuration
+command sbatch --export=PATH
+option name=* --job-name $0
+option time=* --time $0
+option mem=* --mem-per-cpu $0
+option mem=0
+option num_threads=* --cpus-per-task $0
+option num_threads=1 --cpus-per-task 1
+option num_nodes=* --nodes $0
+default gpu=0
+option gpu=0 -p cpu
+option gpu=* -p gpu --gres=gpu:$0 -c $0  # Recommend allocating more CPU than, or equal to the number of GPU
+# note: the --max-jobs-run option is supported as a special case
+# by slurm.pl and you don't have to handle it in the config file.
diff --git a/egs2/mr_openslr64/asr1/conf/train_asr.yaml b/egs2/mr_openslr64/asr1/conf/train_asr.yaml
new file mode 120000
index 00000000000..10fea4f0971
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/train_asr.yaml
@@ -0,0 +1 @@
+tuning/train_asr_conformer.yaml
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/conf/train_lm.yaml b/egs2/mr_openslr64/asr1/conf/train_lm.yaml
new file mode 100644
index 00000000000..5bc7b7fcbfc
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/train_lm.yaml
@@ -0,0 +1,14 @@
+lm_conf:
+    nlayers: 2
+    unit: 650
+optim: sgd        # or adam
+batch_type: folded
+batch_size: 64   # batch size in LM training
+max_epoch: 20      # if the data size is large, we can reduce this
+patience: 3
+
+best_model_criterion:
+-   - valid
+    - loss
+    - min
+keep_nbest_models: 1
diff --git a/egs2/mr_openslr64/asr1/conf/tuning/decode_rnn.yaml b/egs2/mr_openslr64/asr1/conf/tuning/decode_rnn.yaml
new file mode 100644
index 00000000000..169f05a1ad9
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/tuning/decode_rnn.yaml
@@ -0,0 +1,6 @@
+lm_weight: 0.3
+beam_size: 20
+penalty: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+ctc_weight: 0.6
diff --git a/egs2/mr_openslr64/asr1/conf/tuning/decode_transformer.yaml b/egs2/mr_openslr64/asr1/conf/tuning/decode_transformer.yaml
new file mode 100644
index 00000000000..d8671a16988
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/tuning/decode_transformer.yaml
@@ -0,0 +1,7 @@
+batch_size: 16
+beam_size: 20
+penalty: 0.0
+maxlenratio: 0.0
+minlenratio: 0.0
+ctc_weight: 0.5
+lm_weight: 0.3
diff --git a/egs2/mr_openslr64/asr1/conf/tuning/train_asr_conformer.yaml b/egs2/mr_openslr64/asr1/conf/tuning/train_asr_conformer.yaml
new file mode 100644
index 00000000000..f47a0df534c
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/tuning/train_asr_conformer.yaml
@@ -0,0 +1,67 @@
+batch_type: numel
+batch_bins: 10000
+accum_grad: 3
+max_epoch: 60
+patience: none
+init: xavier_uniform
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 5
+
+encoder: conformer
+encoder_conf:
+    output_size: 512
+    attention_heads: 4
+    linear_units: 1024
+    num_blocks: 3
+    dropout_rate: 0.3
+    positional_dropout_rate: 0.3
+    attention_dropout_rate: 0.3
+    input_layer: conv2d
+    normalize_before: true
+    macaron_style: false
+    pos_enc_layer_type: "rel_pos"
+    selfattention_layer_type: "rel_selfattn"
+    activation_type: "swish"
+    use_cnn_module:  true
+    cnn_module_kernel: 17
+
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 1024
+    num_blocks: 3
+    dropout_rate: 0.3
+    positional_dropout_rate: 0.3
+    self_attention_dropout_rate: 0.3
+    src_attention_dropout_rate: 0.3
+
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1
+    length_normalized_loss: false
+
+optim: adam
+optim_conf:
+    lr: 0.0005
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 20000
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 40
+    num_time_mask: 2
diff --git a/egs2/mr_openslr64/asr1/conf/tuning/train_asr_conformer_xlsr.yaml b/egs2/mr_openslr64/asr1/conf/tuning/train_asr_conformer_xlsr.yaml
new file mode 100644
index 00000000000..1dbd14da380
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/tuning/train_asr_conformer_xlsr.yaml
@@ -0,0 +1,88 @@
+batch_type: numel
+batch_bins: 10000
+accum_grad: 3
+max_epoch: 60
+patience: none
+init: xavier_uniform
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 5
+
+freeze_param: [
+"frontend.upstream"
+]
+
+frontend_conf:
+  n_fft: 512
+  hop_length: 256
+
+frontend: s3prl
+frontend_conf:
+  frontend_conf:
+    upstream: wav2vec2_xlsr  # Note: If the upstream is changed, please change the input_size in the preencoder.
+  download_dir: ./hub
+  multilayer_feature: True
+
+preencoder: linear
+preencoder_conf:
+    input_size: 1024  # Note: If the upstream is changed, please change this value accordingly.
+    output_size: 80
+
+encoder: conformer
+encoder_conf:
+    output_size: 512
+    attention_heads: 4
+    linear_units: 1024
+    num_blocks: 3
+    dropout_rate: 0.3
+    positional_dropout_rate: 0.3
+    attention_dropout_rate: 0.3
+    input_layer: conv2d
+    normalize_before: true
+    macaron_style: false
+    pos_enc_layer_type: "rel_pos"
+    selfattention_layer_type: "rel_selfattn"
+    activation_type: "swish"
+    use_cnn_module:  true
+    cnn_module_kernel: 17
+
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 1024
+    num_blocks: 3
+    dropout_rate: 0.3
+    positional_dropout_rate: 0.3
+    self_attention_dropout_rate: 0.3
+    src_attention_dropout_rate: 0.3
+
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1
+    length_normalized_loss: false
+    extract_feats_in_collect_stats: false   # Note: "False" means during collect stats (stage 10), generating dummy stats files rather than extract_feats by forward frontend.
+
+optim: adam
+optim_conf:
+    lr: 0.0005
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 20000
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 40
+    num_time_mask: 2
diff --git a/egs2/mr_openslr64/asr1/conf/tuning/train_asr_rnn.yaml b/egs2/mr_openslr64/asr1/conf/tuning/train_asr_rnn.yaml
new file mode 100644
index 00000000000..f55ad9d72bd
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/tuning/train_asr_rnn.yaml
@@ -0,0 +1,35 @@
+# network architecture
+# encoder related
+encoder: vgg_rnn
+encoder_conf:
+    rnn_type: lstm     # encoder architecture type
+    bidirectional: True
+    use_projection: True
+    num_layers: 4
+    hidden_size: 1024
+    output_size: 1024
+
+# decoder related
+decoder: rnn
+decoder_conf:
+    num_layers: 2
+    hidden_size: 1024
+    sampling_probability: 0
+    att_conf:
+        atype: location
+        adim: 1024
+        aconv_chans: 10
+        aconv_filts: 100
+
+# hybrid CTC/attention
+model_conf:
+    ctc_weight: 0.5
+
+# minibatch related
+batch_size: 30
+
+# optimization related
+optim: adadelta
+max_epoch: 15
+patience: 3
+
diff --git a/egs2/mr_openslr64/asr1/conf/tuning/train_asr_transformer.yaml b/egs2/mr_openslr64/asr1/conf/tuning/train_asr_transformer.yaml
new file mode 100644
index 00000000000..a0a37a5c0e4
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/conf/tuning/train_asr_transformer.yaml
@@ -0,0 +1,68 @@
+# This configuration requires 4 GPUs with 32GB memory
+batch_type: numel
+batch_bins: 10000
+accum_grad: 3
+max_epoch: 60
+patience: none
+init: xavier_uniform
+best_model_criterion:
+-   - valid
+    - acc
+    - max
+keep_nbest_models: 5
+
+encoder: conformer
+encoder_conf:
+    output_size: 512
+    attention_heads: 4
+    linear_units: 1024
+    num_blocks: 3
+    dropout_rate: 0.3
+    positional_dropout_rate: 0.3
+    attention_dropout_rate: 0.3
+    input_layer: conv2d
+    normalize_before: true
+    macaron_style: false
+    pos_enc_layer_type: "rel_pos"
+    selfattention_layer_type: "rel_selfattn"
+    activation_type: "swish"
+    use_cnn_module:  true
+    cnn_module_kernel: 17
+
+decoder: transformer
+decoder_conf:
+    attention_heads: 4
+    linear_units: 1024
+    num_blocks: 3
+    dropout_rate: 0.3
+    positional_dropout_rate: 0.3
+    self_attention_dropout_rate: 0.3
+    src_attention_dropout_rate: 0.3
+
+model_conf:
+    ctc_weight: 0.3
+    lsm_weight: 0.1
+    length_normalized_loss: false
+
+optim: adam
+optim_conf:
+    lr: 0.0005
+scheduler: warmuplr
+scheduler_conf:
+    warmup_steps: 20000
+
+specaug: specaug
+specaug_conf:
+    apply_time_warp: true
+    time_warp_window: 5
+    time_warp_mode: bicubic
+    apply_freq_mask: true
+    freq_mask_width_range:
+    - 0
+    - 30
+    num_freq_mask: 2
+    apply_time_mask: true
+    time_mask_width_range:
+    - 0
+    - 40
+    num_time_mask: 2
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/db.sh b/egs2/mr_openslr64/asr1/db.sh
new file mode 120000
index 00000000000..50d86130898
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/db.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/db.sh
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/local/data.sh b/egs2/mr_openslr64/asr1/local/data.sh
new file mode 100755
index 00000000000..c296b907d59
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/local/data.sh
@@ -0,0 +1,63 @@
+#!/bin/bash
+
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+. ./path.sh || exit 1;
+. ./cmd.sh || exit 1;
+. ./db.sh || exit 1;
+
+# general configuration
+stage=0       # start from 0 if you need to start from data preparation
+stop_stage=1
+# inclusive, was 100
+SECONDS=0
+
+log() {
+    local fname=${BASH_SOURCE[1]##*/}
+    echo -e "$(date '+%Y-%m-%dT%H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+. utils/parse_options.sh
+
+log "data preparation started"
+
+if [ -z "$MARATHI" ]; then
+    log "Variable MARATHI not set in db.sh"
+    exit 2
+fi
+
+mkdir -p ${MARATHI}
+
+workspace=$PWD
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+    log "sub-stage 0: Download Data to downloads"
+
+    cd ${MARATHI}
+    wget https://www.openslr.org/resources/64/mr_in_female.zip
+    unzip -o mr_in_female.zip
+    rm -f mr_in_female.zip
+    
+    cd $workspace    
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+    log "sub-stage 1: Preparing Data for openslr"
+
+    python3 local/data_prep.py -d ${MARATHI}
+    utils/spk2utt_to_utt2spk.pl data/marathi_train/spk2utt > data/marathi_train/utt2spk
+    utils/spk2utt_to_utt2spk.pl data/marathi_dev/spk2utt > data/marathi_dev/utt2spk
+    utils/spk2utt_to_utt2spk.pl data/marathi_test/spk2utt > data/marathi_test/utt2spk
+    utils/fix_data_dir.sh data/marathi_train
+    utils/fix_data_dir.sh data/marathi_dev
+    utils/fix_data_dir.sh data/marathi_test
+fi
+
+log "Successfully finished. [elapsed=${SECONDS}s]"
diff --git a/egs2/mr_openslr64/asr1/local/data_prep.py b/egs2/mr_openslr64/asr1/local/data_prep.py
new file mode 100644
index 00000000000..ed446ef71ae
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/local/data_prep.py
@@ -0,0 +1,99 @@
+#!/usr/bin/env python3
+
+# Copyright 2021 Carnegie Mellon University (Peter Wu)
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+
+import argparse
+import os
+import random
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-d", help="downloads directory", type=str, default="downloads")
+    args = parser.parse_args()
+
+    tsv_path = "%s/line_index.tsv" % args.d
+
+    with open(tsv_path, "r") as inf:
+        tsv_lines = inf.readlines()
+    tsv_lines = [line.strip() for line in tsv_lines]
+
+    spk2utt = {}
+    utt2text = {}
+    for line in tsv_lines:
+        l_list = line.split("\t")
+        fid = l_list[0]
+        spk = l_list[0].split("_")[1]
+        text = l_list[1]
+        path = "%s/%s.wav" % (args.d, fid)
+        if os.path.exists(path):
+            utt2text[fid] = text
+            if spk in spk2utt:
+                spk2utt[spk].append(fid)
+            else:
+                spk2utt[spk] = [fid]
+
+    spks = sorted(list(spk2utt.keys()))
+    num_fids = 0
+    num_test_spks = 0
+    for spk in spks:
+        num_test_spks += 1
+        fids = sorted(list(set(spk2utt[spk])))
+        num_fids += len(fids)
+        if num_fids >= 2000:
+            break
+
+    num_test_spks = 2
+    test_spks = spks[:num_test_spks]
+    train_dev_spks = spks[num_test_spks:]
+    random.Random(0).shuffle(train_dev_spks)
+    num_train = int(len(train_dev_spks) * 0.9)
+    train_spks = train_dev_spks[:num_train]
+    dev_spks = train_dev_spks[num_train:]
+
+    spks_by_phase = {"train": train_spks, "dev": dev_spks, "test": test_spks}
+    flac_dir = "%s" % args.d
+    sr = 16000
+    for phase in spks_by_phase:
+        spks = spks_by_phase[phase]
+        text_strs = []
+        wav_scp_strs = []
+        spk2utt_strs = []
+        num_fids = 0
+        for spk in spks:
+            fids = sorted(list(set(spk2utt[spk])))
+            num_fids += len(fids)
+            if phase == "test" and num_fids > 2000:
+                curr_num_fids = num_fids - 2000
+                random.Random(1).shuffle(fids)
+                fids = fids[:curr_num_fids]
+            utts = [spk + "-" + f for f in fids]
+            utts_str = " ".join(utts)
+            spk2utt_strs.append("%s %s" % (spk, utts_str))
+            for fid, utt in zip(fids, utts):
+                cmd = "ffmpeg -i %s/%s.wav -f wav -ar %d -ab 16 -ac 1 - |" % (
+                    flac_dir,
+                    fid,
+                    sr,
+                )
+                text_strs.append("%s %s" % (utt, utt2text[fid]))
+                wav_scp_strs.append("%s %s" % (utt, cmd))
+        phase_dir = "data/marathi_%s" % phase
+        if not os.path.exists(phase_dir):
+            os.makedirs(phase_dir)
+
+        text_strs = sorted(text_strs)
+        wav_scp_strs = sorted(wav_scp_strs)
+        spk2utt_strs = sorted(spk2utt_strs)
+
+        with open(os.path.join(phase_dir, "text"), "w+") as ouf:
+            for s in text_strs:
+                ouf.write("%s\n" % s)
+        with open(os.path.join(phase_dir, "wav.scp"), "w+") as ouf:
+            for s in wav_scp_strs:
+                ouf.write("%s\n" % s)
+        with open(os.path.join(phase_dir, "spk2utt"), "w+") as ouf:
+            for s in spk2utt_strs:
+                ouf.write("%s\n" % s)
diff --git a/egs2/mr_openslr64/asr1/local/path.sh b/egs2/mr_openslr64/asr1/local/path.sh
new file mode 100644
index 00000000000..e69de29bb2d
diff --git a/egs2/mr_openslr64/asr1/path.sh b/egs2/mr_openslr64/asr1/path.sh
new file mode 120000
index 00000000000..c9ac0a75bc6
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/path.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/path.sh
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/pyscripts b/egs2/mr_openslr64/asr1/pyscripts
new file mode 120000
index 00000000000..ac68ad75b60
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/pyscripts
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/pyscripts
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/run.sh b/egs2/mr_openslr64/asr1/run.sh
new file mode 100755
index 00000000000..4b3fced2fb5
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/run.sh
@@ -0,0 +1,41 @@
+#!/bin/bash
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+lid=false # whether to use language id as additional label
+
+train_set="marathi_train"
+train_dev="marathi_dev"
+test_set="marathi_test"
+
+asr_config=conf/tuning/train_asr_conformer_xlsr.yaml
+inference_config=conf/decode_asr.yaml
+
+ngpu=1
+
+./asr.sh \
+    --stage 1 \
+    --stop_stage 100 \
+    --ngpu ${ngpu} \
+    --nj 10 \
+    --inference_nj 10 \
+    --gpu_inference true \
+    --audio_format "wav" \
+    --inference_args "--batch_size 1" \
+    --use_lm false \
+    --token_type bpe \
+    --nbpe 150 \
+    --feats_type raw \
+    --feats_normalize utt_mvn \
+    --asr_config "${asr_config}" \
+    --inference_config "${inference_config}" \
+    --train_set "${train_set}" \
+    --valid_set "${train_dev}" \
+    --test_sets "${test_set}" \
+    --speed_perturb_factors "0.9 1.0 1.1" \
+    --bpe_train_text "data/${train_set}/text" \
+    --lm_train_text "data/${train_set}/text" \
+    --local_score_opts "--score_lang_id ${lid}" "$@"
diff --git a/egs2/mr_openslr64/asr1/scripts b/egs2/mr_openslr64/asr1/scripts
new file mode 120000
index 00000000000..b25829705dc
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/scripts
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/scripts
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/steps b/egs2/mr_openslr64/asr1/steps
new file mode 120000
index 00000000000..69ab7056139
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/steps
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/steps
\ No newline at end of file
diff --git a/egs2/mr_openslr64/asr1/utils b/egs2/mr_openslr64/asr1/utils
new file mode 120000
index 00000000000..e18ae14b549
--- /dev/null
+++ b/egs2/mr_openslr64/asr1/utils
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/utils
\ No newline at end of file

From aee683f1559fe922236a6c4935c0485652e4ea3d Mon Sep 17 00:00:00 2001
From: simpleoier <netnetchangxk@gmail.com>
Date: Fri, 1 Apr 2022 09:01:13 -0400
Subject: [PATCH 22/22] fix issues about longformer package upgrades pytorch
 package. Because pytorch-lightning requires torch>=1.8.*

---
 .../asr1/local/check_train_test_duplicate.py           |  2 +-
 tools/installers/install_longformer.sh                 | 10 +++++-----
 2 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py b/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py
index 4a2720e777f..3c61d786c1d 100644
--- a/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py
+++ b/egs2/zh_openslr38/asr1/local/check_train_test_duplicate.py
@@ -32,7 +32,7 @@
     print(count, "duplicates in", test_name)
 
     # if input("continue? [y/n]") == 'y':
-    # remove all instances of duplicate uttids in these files: spk2utt, text, utt2spk, wav.scp
+    # remove all instances of duplicate uttids in: spk2utt, text, utt2spk, wav.scp
     with open(f"data/{test_name}/spk2utt", "r") as f:
         # replace all uttid with empty string
         text = f.read()
diff --git a/tools/installers/install_longformer.sh b/tools/installers/install_longformer.sh
index a2f11eecf15..c942abb0dd9 100755
--- a/tools/installers/install_longformer.sh
+++ b/tools/installers/install_longformer.sh
@@ -36,12 +36,12 @@ if ! "${python_36_plus}"; then
     exit 1
 else
 
-    if $(pt_plus 1.6.1); then
-        pip install git+https://github.com/roshansh-cmu/longformer.git
-        pip install datasets bert-score
-	pip install git+https://github.com/Maluuba/nlg-eval.git@master 
+    if $(pt_plus 1.8.0); then
+        python -m pip install git+https://github.com/roshansh-cmu/longformer.git
+        python -m pip install datasets bert-score
+        python -m pip install git+https://github.com/Maluuba/nlg-eval.git@master
     else
-        echo "[WARNING] Longformer requires pytorch>=1.6.1"
+        echo "[WARNING] Longformer requires pytorch>=1.8.*"
     fi
 
 fi