Merge branch 'espnet:master' into master

chintu619 · Apr 6, 2022 · fcf13c4 · fcf13c4
2 parents 39700a0 + c4aba12
commit fcf13c4
Show file tree

Hide file tree

Showing 60 changed files with 1,387 additions and 27 deletions.
diff --git a/egs2/README.md b/egs2/README.md
@@ -57,6 +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/                                                                                   |              |
+| 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         |                                                                                                             |              |
@@ -97,3 +98,4 @@ 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                                                                                    |              |
diff --git a/egs2/TEMPLATE/asr1/db.sh b/egs2/TEMPLATE/asr1/db.sh
@@ -128,7 +128,9 @@ PRIMEWORDS_CHINESE=downloads
 SEAME=
 BENGALI=downloads
 IWSLT14=
+ST_CMDS=downloads
 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
@@ -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
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/asr.sh
diff --git a/egs2/mr_openslr64/asr1/cmd.sh b/egs2/mr_openslr64/asr1/cmd.sh
@@ -0,0 +1 @@
+../../TEMPLATE/asr1/cmd.sh
diff --git a/egs2/mr_openslr64/asr1/conf/decode_asr.yaml b/egs2/mr_openslr64/asr1/conf/decode_asr.yaml
@@ -0,0 +1 @@
+tuning/decode_transformer.yaml
diff --git a/egs2/mr_openslr64/asr1/conf/fbank.conf 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
@@ -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
@@ -0,0 +1 @@
+--sample-frequency=16000
diff --git a/egs2/mr_openslr64/asr1/conf/queue.conf 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
@@ -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
@@ -0,0 +1 @@
+tuning/train_asr_conformer.yaml
diff --git a/egs2/mr_openslr64/asr1/conf/train_lm.yaml 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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
+