From b422e7a97f6eaa6baede95a19cd24564ba627879 Mon Sep 17 00:00:00 2001 From: Yuekai Zhang Date: Wed, 6 Mar 2024 22:09:01 +0800 Subject: [PATCH] add speechio --- egs/multi_zh-hans/ASR/prepare.sh | 3 + .../ASR/local/compute_fbank_speechio.py | 131 ++ .../ASR/local/display_manifest_statistics.py | 1159 ++++++++++++++ egs/speechio/ASR/prepare.sh | 67 + egs/speechio/ASR/shared | 1 + egs/speechio/ASR/whisper/asr_datamodule.py | 197 +++ egs/speechio/ASR/whisper/decode.py | 520 +++++++ egs/speechio/ASR/whisper/multi_dataset.py | 61 + egs/speechio/ASR/whisper/requirements.txt | 1 + .../whisper_encoder_forward_monkey_patch.py | 1 + egs/speechio/ASR/zipformer/asr_datamodule.py | 1 + egs/speechio/ASR/zipformer/beam_search.py | 1 + egs/speechio/ASR/zipformer/ctc_decode.py | 623 ++++++++ egs/speechio/ASR/zipformer/decode.py | 828 ++++++++++ egs/speechio/ASR/zipformer/decoder.py | 1 + .../ASR/zipformer/encoder_interface.py | 1 + ...asr-multi-zh-hans-zipformer-ctc-2023-10-24 | 1 + egs/speechio/ASR/zipformer/joiner.py | 1 + egs/speechio/ASR/zipformer/model.py | 1 + egs/speechio/ASR/zipformer/multi_dataset.py | 1 + egs/speechio/ASR/zipformer/optim.py | 1 + egs/speechio/ASR/zipformer/scaling.py | 1 + .../ASR/zipformer/scaling_converter.py | 1 + egs/speechio/ASR/zipformer/subsampling.py | 1 + egs/speechio/ASR/zipformer/train.py | 1385 +++++++++++++++++ egs/speechio/ASR/zipformer/zipformer.py | 1 + 26 files changed, 4990 insertions(+) create mode 100644 egs/speechio/ASR/local/compute_fbank_speechio.py create mode 100644 egs/speechio/ASR/local/display_manifest_statistics.py create mode 100644 egs/speechio/ASR/prepare.sh create mode 120000 egs/speechio/ASR/shared create mode 100644 egs/speechio/ASR/whisper/asr_datamodule.py create mode 100644 egs/speechio/ASR/whisper/decode.py create mode 100644 egs/speechio/ASR/whisper/multi_dataset.py create mode 120000 egs/speechio/ASR/whisper/requirements.txt create mode 120000 egs/speechio/ASR/whisper/whisper_encoder_forward_monkey_patch.py create mode 120000 egs/speechio/ASR/zipformer/asr_datamodule.py create mode 120000 egs/speechio/ASR/zipformer/beam_search.py create mode 100644 egs/speechio/ASR/zipformer/ctc_decode.py create mode 100644 egs/speechio/ASR/zipformer/decode.py create mode 120000 egs/speechio/ASR/zipformer/decoder.py create mode 120000 egs/speechio/ASR/zipformer/encoder_interface.py create mode 160000 egs/speechio/ASR/zipformer/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24 create mode 120000 egs/speechio/ASR/zipformer/joiner.py create mode 120000 egs/speechio/ASR/zipformer/model.py create mode 120000 egs/speechio/ASR/zipformer/multi_dataset.py create mode 120000 egs/speechio/ASR/zipformer/optim.py create mode 120000 egs/speechio/ASR/zipformer/scaling.py create mode 120000 egs/speechio/ASR/zipformer/scaling_converter.py create mode 120000 egs/speechio/ASR/zipformer/subsampling.py create mode 100644 egs/speechio/ASR/zipformer/train.py create mode 120000 egs/speechio/ASR/zipformer/zipformer.py diff --git a/egs/multi_zh-hans/ASR/prepare.sh b/egs/multi_zh-hans/ASR/prepare.sh index d9fce3251e..96ae1cf60a 100755 --- a/egs/multi_zh-hans/ASR/prepare.sh +++ b/egs/multi_zh-hans/ASR/prepare.sh @@ -107,6 +107,9 @@ if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then if [ -e ../../aishell4/ASR/data/fbank/.fbank.done ]; then cd data/fbank ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_test) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train_L) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train_M) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train_S) . ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_L.jsonl.gz) . ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_M.jsonl.gz) . ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_S.jsonl.gz) . diff --git a/egs/speechio/ASR/local/compute_fbank_speechio.py b/egs/speechio/ASR/local/compute_fbank_speechio.py new file mode 100644 index 0000000000..d6956781b4 --- /dev/null +++ b/egs/speechio/ASR/local/compute_fbank_speechio.py @@ -0,0 +1,131 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Fangjun Kuang +# Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the ST-CMDS dataset. +It looks for manifests in the directory data/manifests/stcmds. + +The generated fbank features are saved in data/fbank. +""" + +import argparse +import logging +import os +from pathlib import Path + +import torch +from lhotse import CutSet, WhisperFbank, WhisperFbankConfig, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor, str2bool + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +SPEECHIO_TESTSET_INDEX = 26 # Currently, from 0 - 26 test sets are open source. + +def compute_fbank_speechio(num_mel_bins: int = 80, speed_perturb: bool = False, fbank_dir: str = "data/fbank", whisper_fbank: bool = False): + src_dir = Path("data/manifests") + output_dir = Path(fbank_dir) + num_jobs = min(8, os.cpu_count()) + + dataset_parts = [] + for i in range(SPEECHIO_TESTSET_INDEX + 1): + idx = f"{i}".zfill(2) + dataset_parts.append(f"SPEECHIO_ASR_ZH000{idx}") + + prefix = "speechio" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + if whisper_fbank: + extractor = WhisperFbank(WhisperFbankConfig(num_filters=args.num_mel_bins, device='cuda')) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}") + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) + parser.add_argument( + "--fbank-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + return parser.parse_args() + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + args = get_args() + compute_fbank_speechio( + num_mel_bins=args.num_mel_bins, fbank_dir=args.fbank_dir, whisper_fbank=args.whisper_fbank + ) diff --git a/egs/speechio/ASR/local/display_manifest_statistics.py b/egs/speechio/ASR/local/display_manifest_statistics.py new file mode 100644 index 0000000000..b2f52d137e --- /dev/null +++ b/egs/speechio/ASR/local/display_manifest_statistics.py @@ -0,0 +1,1159 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +This file displays duration statistics of utterances in a manifest. +You can use the displayed value to choose minimum/maximum duration +to remove short and long utterances during the training. + +See the function `remove_short_and_long_utt()` in transducer_stateless/train.py +for usage. +""" + +SPEECHIO_TESTSET_INDEX = 26 # Currently, from 0 - 26 test sets are open source. + +from lhotse import load_manifest_lazy + + +def main(): + dataset_parts = [] + for i in range(SPEECHIO_TESTSET_INDEX + 1): + idx = f"{i}".zfill(2) + dataset_parts.append(f"SPEECHIO_ASR_ZH000{idx}") + + prefix="speechio" + suffix="jsonl.gz" + + for partition in dataset_parts: + path = f"./data/fbank/{prefix}_cuts_{partition}.{suffix}" + cuts = load_manifest_lazy(path) + print(f"===================Duration statistics of {partition}===================") + cuts.describe() + +if __name__ == "__main__": + main() + +""" +===================Duration statistics of SPEECHIO_ASR_ZH00000=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 879 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:36:09 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.6 │ +├───────────────────────────┼──────────┤ +│ std │ 2.0 │ +├───────────────────────────┼──────────┤ +│ min │ 1.7 │ +├───────────────────────────┼──────────┤ +│ 25% │ 5.0 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.5 │ +├───────────────────────────┼──────────┤ +│ 75% │ 8.1 │ +├───────────────────────────┼──────────┤ +│ 99% │ 11.2 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 11.6 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.2 │ +├───────────────────────────┼──────────┤ +│ max │ 12.5 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 879 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 879 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 879 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:36:09 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:36:09 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00001=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 5069 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 08:43:04 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.2 │ +├───────────────────────────┼──────────┤ +│ std │ 2.1 │ +├───────────────────────────┼──────────┤ +│ min │ 0.6 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.6 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.2 │ +├───────────────────────────┼──────────┤ +│ 75% │ 7.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.7 │ +├───────────────────────────┼──────────┤ +│ max │ 12.5 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 5069 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 5069 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 5069 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 08:43:04 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 08:43:04 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00002=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 2993 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:45:09 │ +├───────────────────────────┼──────────┤ +│ mean │ 3.3 │ +├───────────────────────────┼──────────┤ +│ std │ 1.5 │ +├───────────────────────────┼──────────┤ +│ min │ 0.4 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.2 │ +├───────────────────────────┼──────────┤ +│ 50% │ 3.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 4.3 │ +├───────────────────────────┼──────────┤ +│ 99% │ 7.3 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 7.8 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 9.1 │ +├───────────────────────────┼──────────┤ +│ max │ 11.8 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 2993 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 2993 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 2993 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:45:09 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:45:09 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00003=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1683 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:23:28 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.1 │ +├───────────────────────────┼──────────┤ +│ std │ 1.4 │ +├───────────────────────────┼──────────┤ +│ min │ 2.4 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.0 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.0 │ +├───────────────────────────┼──────────┤ +│ 99% │ 9.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.4 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.8 │ +├───────────────────────────┼──────────┤ +│ max │ 14.2 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1683 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1683 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1683 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:23:28 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:23:28 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00004=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1311 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:49:16 │ +├───────────────────────────┼──────────┤ +│ mean │ 7.7 │ +├───────────────────────────┼──────────┤ +│ std │ 2.8 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 5.8 │ +├───────────────────────────┼──────────┤ +│ 50% │ 8.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 9.8 │ +├───────────────────────────┼──────────┤ +│ 99% │ 12.9 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 13.5 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 13.8 │ +├───────────────────────────┼──────────┤ +│ max │ 14.4 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1311 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1311 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1311 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:49:16 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:49:16 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00005=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 3148 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 04:22:47 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.0 │ +├───────────────────────────┼──────────┤ +│ std │ 1.4 │ +├───────────────────────────┼──────────┤ +│ min │ 2.0 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.9 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 8.8 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.3 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.3 │ +├───────────────────────────┼──────────┤ +│ max │ 11.1 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 3148 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 3148 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 3148 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 04:22:47 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 04:22:47 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00006=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1561 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:39:33 │ +├───────────────────────────┼──────────┤ +│ mean │ 3.8 │ +├───────────────────────────┼──────────┤ +│ std │ 2.2 │ +├───────────────────────────┼──────────┤ +│ min │ 0.4 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.2 │ +├───────────────────────────┼──────────┤ +│ 50% │ 3.3 │ +├───────────────────────────┼──────────┤ +│ 75% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.4 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 11.3 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.3 │ +├───────────────────────────┼──────────┤ +│ max │ 23.8 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1561 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1561 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1561 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:39:33 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:39:33 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00007=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 770 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 00:58:57 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.6 │ +├───────────────────────────┼──────────┤ +│ std │ 2.4 │ +├───────────────────────────┼──────────┤ +│ min │ 0.7 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.7 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.0 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.0 │ +├───────────────────────────┼──────────┤ +│ 99% │ 11.8 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 13.0 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.1 │ +├───────────────────────────┼──────────┤ +│ max │ 18.7 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 770 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 770 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 770 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 00:58:57 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 00:58:57 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00008=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 884 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:16:55 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.2 │ +├───────────────────────────┼──────────┤ +│ std │ 2.3 │ +├───────────────────────────┼──────────┤ +│ min │ 1.1 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.5 │ +├───────────────────────────┼──────────┤ +│ 50% │ 5.0 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.4 │ +├───────────────────────────┼──────────┤ +│ 99% │ 11.3 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 12.7 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 16.2 │ +├───────────────────────────┼──────────┤ +│ max │ 18.5 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 884 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 884 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 884 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:16:55 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:16:55 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00009=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 3466 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 04:38:13 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.8 │ +├───────────────────────────┼──────────┤ +│ std │ 1.9 │ +├───────────────────────────┼──────────┤ +│ min │ 1.1 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.4 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.5 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.5 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 11.3 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.5 │ +├───────────────────────────┼──────────┤ +│ max │ 13.1 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 3466 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 3466 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 3466 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 04:38:13 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 04:38:13 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00010=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 2251 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 04:12:54 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.7 │ +├───────────────────────────┼──────────┤ +│ std │ 3.0 │ +├───────────────────────────┼──────────┤ +│ min │ 1.4 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.5 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.3 │ +├───────────────────────────┼──────────┤ +│ 75% │ 8.5 │ +├───────────────────────────┼──────────┤ +│ 99% │ 14.9 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 15.5 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.8 │ +├───────────────────────────┼──────────┤ +│ max │ 16.2 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 2251 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 2251 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 2251 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 04:12:54 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 04:12:54 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00011=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1053 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 03:27:12 │ +├───────────────────────────┼──────────┤ +│ mean │ 11.8 │ +├───────────────────────────┼──────────┤ +│ std │ 3.4 │ +├───────────────────────────┼──────────┤ +│ min │ 1.1 │ +├───────────────────────────┼──────────┤ +│ 25% │ 11.5 │ +├───────────────────────────┼──────────┤ +│ 50% │ 13.0 │ +├───────────────────────────┼──────────┤ +│ 75% │ 13.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 15.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 15.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 20.7 │ +├───────────────────────────┼──────────┤ +│ max │ 22.2 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1053 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1053 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1053 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 03:27:12 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 03:27:12 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00012=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1170 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 03:23:34 │ +├───────────────────────────┼──────────┤ +│ mean │ 10.4 │ +├───────────────────────────┼──────────┤ +│ std │ 3.5 │ +├───────────────────────────┼──────────┤ +│ min │ 0.8 │ +├───────────────────────────┼──────────┤ +│ 25% │ 8.0 │ +├───────────────────────────┼──────────┤ +│ 50% │ 11.5 │ +├───────────────────────────┼──────────┤ +│ 75% │ 13.2 │ +├───────────────────────────┼──────────┤ +│ 99% │ 15.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 15.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.7 │ +├───────────────────────────┼──────────┤ +│ max │ 20.3 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1170 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1170 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1170 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 03:23:34 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 03:23:34 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00013=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1321 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:46:41 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.8 │ +├───────────────────────────┼──────────┤ +│ std │ 1.5 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.8 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.8 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.8 │ +├───────────────────────────┼──────────┤ +│ 99% │ 8.5 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 9.5 │ +├───────────────────────────┼──────────┤ +│ max │ 9.7 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1321 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1321 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1321 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:46:41 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:46:41 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00014=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 856 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:00:39 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.3 │ +├───────────────────────────┼──────────┤ +│ std │ 1.8 │ +├───────────────────────────┼──────────┤ +│ min │ 0.8 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.9 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.5 │ +├───────────────────────────┼──────────┤ +│ 99% │ 8.5 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.2 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ max │ 11.1 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 856 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 856 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 856 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:00:39 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:00:39 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00015=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1168 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:08:52 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.6 │ +├───────────────────────────┼──────────┤ +│ std │ 2.0 │ +├───────────────────────────┼──────────┤ +│ min │ 1.2 │ +├───────────────────────────┼──────────┤ +│ 25% │ 5.3 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.8 │ +├───────────────────────────┼──────────┤ +│ 75% │ 8.2 │ +├───────────────────────────┼──────────┤ +│ 99% │ 9.9 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.1 │ +├───────────────────────────┼──────────┤ +│ max │ 15.5 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1168 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1168 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1168 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:08:52 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:08:52 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00016=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1201 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:00:46 │ +├───────────────────────────┼──────────┤ +│ mean │ 3.0 │ +├───────────────────────────┼──────────┤ +│ std │ 2.0 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 1.6 │ +├───────────────────────────┼──────────┤ +│ 50% │ 2.3 │ +├───────────────────────────┼──────────┤ +│ 75% │ 3.8 │ +├───────────────────────────┼──────────┤ +│ 99% │ 9.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.5 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 9.7 │ +├───────────────────────────┼──────────┤ +│ max │ 9.9 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1201 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1201 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1201 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:00:46 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:00:46 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00017=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1271 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:47:57 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.1 │ +├───────────────────────────┼──────────┤ +│ std │ 2.2 │ +├───────────────────────────┼──────────┤ +│ min │ 1.0 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.3 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.8 │ +├───────────────────────────┼──────────┤ +│ 99% │ 9.7 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ max │ 10.4 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1271 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1271 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1271 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:47:57 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:47:57 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00018=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 899 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 00:51:12 │ +├───────────────────────────┼──────────┤ +│ mean │ 3.4 │ +├───────────────────────────┼──────────┤ +│ std │ 1.2 │ +├───────────────────────────┼──────────┤ +│ min │ 1.3 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.5 │ +├───────────────────────────┼──────────┤ +│ 50% │ 3.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 4.1 │ +├───────────────────────────┼──────────┤ +│ 99% │ 6.7 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 7.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 8.2 │ +├───────────────────────────┼──────────┤ +│ max │ 9.2 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 899 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 899 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 899 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 00:51:12 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 00:51:12 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00019=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 615 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 00:41:43 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.1 │ +├───────────────────────────┼──────────┤ +│ std │ 1.5 │ +├───────────────────────────┼──────────┤ +│ min │ 1.3 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.8 │ +├───────────────────────────┼──────────┤ +│ 50% │ 3.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.2 │ +├───────────────────────────┼──────────┤ +│ 99% │ 7.9 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 8.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 8.6 │ +├───────────────────────────┼──────────┤ +│ max │ 8.8 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 615 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 615 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 615 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 00:41:43 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 00:41:43 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00020=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1590 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:10:54 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.9 │ +├───────────────────────────┼──────────┤ +│ std │ 1.5 │ +├───────────────────────────┼──────────┤ +│ min │ 1.2 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.8 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.0 │ +├───────────────────────────┼──────────┤ +│ 99% │ 8.5 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 8.7 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 9.2 │ +├───────────────────────────┼──────────┤ +│ max │ 10.4 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1590 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1590 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1590 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:10:54 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:10:54 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00021=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1035 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:44:07 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.0 │ +├───────────────────────────┼──────────┤ +│ std │ 1.8 │ +├───────────────────────────┼──────────┤ +│ min │ 1.1 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.7 │ +├───────────────────────────┼──────────┤ +│ 50% │ 5.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 7.3 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.4 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.6 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 11.0 │ +├───────────────────────────┼──────────┤ +│ max │ 11.1 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1035 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1035 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1035 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:44:07 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:44:07 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00022=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1026 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:40:43 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.9 │ +├───────────────────────────┼──────────┤ +│ std │ 2.2 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.4 │ +├───────────────────────────┼──────────┤ +│ 50% │ 5.8 │ +├───────────────────────────┼──────────┤ +│ 75% │ 7.1 │ +├───────────────────────────┼──────────┤ +│ 99% │ 12.1 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 12.7 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 13.9 │ +├───────────────────────────┼──────────┤ +│ max │ 14.0 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1026 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1026 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1026 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:40:43 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:40:43 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00023=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1528 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:06:51 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.0 │ +├───────────────────────────┼──────────┤ +│ std │ 2.5 │ +├───────────────────────────┼──────────┤ +│ min │ 0.5 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.1 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.5 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.6 │ +├───────────────────────────┼──────────┤ +│ 99% │ 12.3 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 13.9 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.8 │ +├───────────────────────────┼──────────┤ +│ max │ 16.8 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1528 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1528 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1528 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:06:51 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:06:51 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00024=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1930 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:39:02 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.9 │ +├───────────────────────────┼──────────┤ +│ std │ 2.0 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.4 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.7 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.2 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.3 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.9 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.0 │ +├───────────────────────────┼──────────┤ +│ max │ 12.6 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1930 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1930 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1930 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:39:02 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:39:02 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00025=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1164 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:24:42 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.4 │ +├───────────────────────────┼──────────┤ +│ std │ 1.9 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.9 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.6 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.4 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.9 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.5 │ +├───────────────────────────┼──────────┤ +│ max │ 13.0 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1164 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1164 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1164 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:24:42 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:24:42 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00026=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1336 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:25:38 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.5 │ +├───────────────────────────┼──────────┤ +│ std │ 2.3 │ +├───────────────────────────┼──────────┤ +│ min │ 0.5 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.8 │ +├───────────────────────────┼──────────┤ +│ 75% │ 8.3 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.4 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 11.9 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.9 │ +├───────────────────────────┼──────────┤ +│ max │ 13.3 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1336 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1336 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1336 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:25:38 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:25:38 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ + +""" diff --git a/egs/speechio/ASR/prepare.sh b/egs/speechio/ASR/prepare.sh new file mode 100644 index 0000000000..5b29440e57 --- /dev/null +++ b/egs/speechio/ASR/prepare.sh @@ -0,0 +1,67 @@ +#!/usr/bin/env bash + +set -eou pipefail + +stage=3 +stop_stage=3 + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/SPEECHIO_ASR_ZH00000 +# This directory contains the following files downloaded from +# https://github.com/SpeechColab/Leaderboard +# +# - metadata.tsv +# - wav +# - wav.scp +# - trans.txt +# + +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +log "dl_dir: $dl_dir" + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare speechio manifest" + # We assume that you have downloaded the speechio dataset + # to $dl_dir + mkdir -p data/manifests + if [ ! -e data/manifests/.speechio.done ]; then + lhotse prepare speechio $dl_dir data/manifests + touch data/manifests/.speechio.done + fi +fi + +whisper_mel_bins=80 +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Compute whisper fbank for speechio" + if [ ! -f data/fbank/.speechio.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_speechio.py --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.speechio.done + fi +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Compute kaldi fbank for speechio" + if [ ! -f data/fbank/.speechio.kaldi.done ]; then + fbank_dir=data/fbank_kaldi + mkdir -p $fbank_dir + ./local/compute_fbank_speechio.py --fbank-dir $fbank_dir + touch data/fbank/.speechio.kaldi.done + fi +fi diff --git a/egs/speechio/ASR/shared b/egs/speechio/ASR/shared new file mode 120000 index 0000000000..9d8803a7d5 --- /dev/null +++ b/egs/speechio/ASR/shared @@ -0,0 +1 @@ +../../../icefall/shared// \ No newline at end of file diff --git a/egs/speechio/ASR/whisper/asr_datamodule.py b/egs/speechio/ASR/whisper/asr_datamodule.py new file mode 100644 index 0000000000..a32ea83e81 --- /dev/null +++ b/egs/speechio/ASR/whisper/asr_datamodule.py @@ -0,0 +1,197 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from lhotse import CutSet, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SimpleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples + AudioSamples, +) +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class AsrDataModule: + """ + DataModule for k2 ASR experiments. + There is no train and valid dataloader, for speechio dataset + but there can be multiple test dataloaders. + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - cut concatenation, + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=300.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--concatenate-cuts", + type=str2bool, + default=False, + help="When enabled, utterances (cuts) will be concatenated " + "to minimize the amount of padding.", + ) + group.add_argument( + "--duration-factor", + type=float, + default=1.0, + help="Determines the maximum duration of a concatenated cut " + "relative to the duration of the longest cut in a batch.", + ) + group.add_argument( + "--gap", + type=float, + default=1.0, + help="The amount of padding (in seconds) inserted between " + "concatenated cuts. This padding is filled with noise when " + "noise augmentation is used.", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + group.add_argument( + "--input-strategy", + type=str, + default="PrecomputedFeatures", + help="AudioSamples or PrecomputedFeatures", + ) + parser.add_argument( + "--start-index", + type=int, + default=0, + help="Decoding will start from dataset SPEECHIO_ASR_ZH000index", + ) + + parser.add_argument( + "--end-index", + type=int, + default=26, + help="Decoding will end with dataset SPEECHIO_ASR_ZH000index", + ) + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=eval(self.args.input_strategy)(), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl diff --git a/egs/speechio/ASR/whisper/decode.py b/egs/speechio/ASR/whisper/decode.py new file mode 100644 index 0000000000..5a0d478eb3 --- /dev/null +++ b/egs/speechio/ASR/whisper/decode.py @@ -0,0 +1,520 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, +# Fangjun Kuang, +# Wei Kang) +# 2024 Yuekai Zhang +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +# Command for decoding using fine-tuned models: +git lfs install +git clone https://huggingface.co/yuekai/icefall_asr_aishell_whisper +ln -s icefall_asr_aishell_whisper/exp_large_v2/epoch-10-avg6.pt whisper/exp_large_v2/epoch-999.pt + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2 \ + --model-name large-v2 \ + --epoch 999 --avg 1 \ + --beam-size 10 --max-duration 50 + +# Command for decoding using pretrained models (before fine-tuning): + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2_pretrained \ + --model-name large-v2 \ + --epoch -1 --avg 1 \ + --start-index 14 --end-index 15 \ + --remove-whisper-encoder-input-length-restriction False \ + --beam-size 1 --max-duration 50 + +""" + +import argparse +import logging +import re +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +import whisper + +from asr_datamodule import AsrDataModule +from tn.chinese.normalizer import Normalizer +from whisper.normalizers import BasicTextNormalizer +from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward +from zhconv import convert +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from icefall.checkpoint import average_checkpoints_with_averaged_model, load_checkpoint +from icefall.env import get_env_info +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def average_checkpoints( + filenames: List[Path], device: torch.device = torch.device("cpu") +) -> dict: + """Average a list of checkpoints. + The function is mainly used for deepspeed converted checkpoint averaging, which only include model state_dict. + + Args: + filenames: + Filenames of the checkpoints to be averaged. We assume all + checkpoints are saved by :func:`save_checkpoint`. + device: + Move checkpoints to this device before averaging. + Returns: + Return a dict (i.e., state_dict) which is the average of all + model state dicts contained in the checkpoints. + """ + n = len(filenames) + + if "model" in torch.load(filenames[0], map_location=device): + avg = torch.load(filenames[0], map_location=device)["model"] + else: + avg = torch.load(filenames[0], map_location=device) + + # Identify shared parameters. Two parameters are said to be shared + # if they have the same data_ptr + uniqued: Dict[int, str] = dict() + + for k, v in avg.items(): + v_data_ptr = v.data_ptr() + if v_data_ptr in uniqued: + continue + uniqued[v_data_ptr] = k + + uniqued_names = list(uniqued.values()) + + for i in range(1, n): + if "model" in torch.load(filenames[i], map_location=device): + state_dict = torch.load(filenames[i], map_location=device)["model"] + else: + state_dict = torch.load(filenames[i], map_location=device) + for k in uniqued_names: + avg[k] += state_dict[k] + + for k in uniqued_names: + if avg[k].is_floating_point(): + avg[k] /= n + else: + avg[k] //= n + + return avg + + +def remove_punctuation(text: str or List[str]): + """Modified from https://github.com/yeyupiaoling/Whisper-Finetune/blob/master/utils/data_utils.py + + Args: + text: It can be a string or a list of strings. + Returns: + Return a string or a list of strings without any punctuation. + """ + punctuation = "!,.;:?、！，。；：？《》 " + if isinstance(text, str): + text = re.sub(r"[{}]+".format(punctuation), "", text).strip() + return text + elif isinstance(text, list): + result_text = [] + for t in text: + t = re.sub(r"[{}]+".format(punctuation), "", t).strip() + result_text.append(t) + return result_text + else: + raise Exception(f"Not support type {type(text)}") + + +def to_simple(text: str or List[str]): + """Convert traditional Chinese to simplified Chinese. + Args: + text: It can be a string or a list of strings. + Returns: + Return a string or a list of strings converted to simplified Chinese. + """ + if isinstance(text, str): + text = convert(text, "zh-cn") + return text + elif isinstance(text, list): + result_text = [] + for t in text: + t = convert(t, "zh-cn") + result_text.append(t) + return result_text + else: + raise Exception(f"Not support type{type(text)}") + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=-1, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=1, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--method", + type=str, + default="beam-search", + help="""Decoding method. + Supported values are: + - beam-search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=1, + help="beam size for beam search decoding", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="whisper/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--model-name", + type=str, + default="large-v2", + choices=["large-v2", "large-v3", "medium", "small", "tiny"], + help="""The model name to use. + """, + ) + + parser.add_argument( + "--remove-whisper-encoder-input-length-restriction", + type=str2bool, + default=True, + help="replace whisper encoder forward method to remove input length restriction", + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + "env_info": get_env_info(), + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + batch: dict, +) -> Dict[str, List[List[int]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: "beam-search" + - value: A list of lists. Each sublist is a list of token IDs. + Args: + params: + It is returned by :func:`get_params`. + model: + The neural model. + batch: + It is returned by :meth:`torch.utils.data.DataLoader.__iter__`. + Returns: + Return a dict, whose key may be "beam-search". + """ + dtype = torch.float16 + device = torch.device("cuda") + + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device, dtype=dtype).transpose(1, 2) + if not params.remove_whisper_encoder_input_length_restriction: + T = 3000 + if feature.shape[2] < T: + feature = torch.cat( + [ + feature, + torch.zeros( + feature.shape[0], feature.shape[1], T - feature.shape[2] + ).to(device, dtype=dtype), + ], + 2, + ) + + supervisions = batch["supervisions"] + feature_len = supervisions["num_frames"] + feature_len = feature_len.to(device, dtype=dtype) + results = model.decode(feature, params.decoding_options) + hyps = [result.text for result in results] + + hyps = remove_punctuation(hyps) + hyps = to_simple(hyps) + hyps = [params.normalizer.normalize(hyp) for hyp in hyps] + print(hyps) + return {"beam-search": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + The dataloader. + params: + It is returned by :func:`get_params`. + model: + The neural model. + Returns: + Return a dict, whose key may be "beam-search". + """ + results = [] + + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + batch=batch, + ) + + for lm_scale, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[lm_scale].extend(this_batch) + + num_cuts += len(batch["supervisions"]["text"]) + + if batch_idx % 100 == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + + enable_log = True + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.exp_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + if enable_log: + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.exp_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + # we compute CER for aishell dataset. + results_char = [] + for res in results: + results_char.append((res[0], list("".join(res[1])), list("".join(res[2])))) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results_char, enable_log=enable_log + ) + test_set_wers[key] = wer + + if enable_log: + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = params.exp_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tCER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + setup_logger( + f"{params.exp_dir}/log-{params.method}-beam{params.beam_size}/log-decode-{params.suffix}" + ) + + options = whisper.DecodingOptions( + task="transcribe", + language="zh", + without_timestamps=True, + beam_size=params.beam_size, + ) + params.decoding_options = options + params.cleaner = BasicTextNormalizer() + params.normalizer = Normalizer() + + logging.info("Decoding started") + logging.info(params) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda") + + logging.info(f"device: {device}") + + if params.remove_whisper_encoder_input_length_restriction: + replace_whisper_encoder_forward() + model = whisper.load_model(params.model_name, "cpu") + if params.epoch > 0: + if params.avg > 1: + start = params.epoch - params.avg + assert start >= 1, start + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu" + ) + if "model" not in checkpoint: + # deepspeed converted checkpoint only contains model state_dict + filenames = [ + f"{params.exp_dir}/epoch-{epoch}.pt" + for epoch in range(start, params.epoch + 1) + ] + model.load_state_dict(average_checkpoints(filenames)) + else: + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + # save checkpoints + filename = f"{params.exp_dir}/epoch-{params.epoch}-avg-{params.avg}.pt" + torch.save(model.state_dict(), filename) + else: + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu" + ) + if "model" not in checkpoint: + model.load_state_dict(checkpoint, strict=True) + else: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + model.to(device) + model.eval() + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir, args.start_index, args.end_index) + + def remove_long_utt(c: Cut): + # Keep only utterances with duration in 30 seconds + # + if c.duration > 30.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + return True + + test_sets_cuts = multi_dataset.test_cuts() + + test_sets = test_sets_cuts.keys() + test_dls = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_long_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + ) + + save_results(params=params, test_set_name=test_set, results_dict=results_dict) + + logging.info("Done!") + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/speechio/ASR/whisper/multi_dataset.py b/egs/speechio/ASR/whisper/multi_dataset.py new file mode 100644 index 0000000000..f427c271f9 --- /dev/null +++ b/egs/speechio/ASR/whisper/multi_dataset.py @@ -0,0 +1,61 @@ +# Copyright 2023 Xiaomi Corp. (authors: Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import glob +import logging +import re +from pathlib import Path +from typing import Dict, List + +import lhotse +from lhotse import CutSet, load_manifest_lazy + + +class MultiDataset: + def __init__(self, fbank_dir: str, start_index: int = 0, end_index: int = 26): + """ + Args: + manifest_dir: + It is expected to contain the following files: + - speechio_cuts_SPEECHIO_ASR_ZH00000.jsonl.gz + ... + - speechio_cuts_SPEECHIO_ASR_ZH00026.jsonl.gz + """ + self.fbank_dir = Path(fbank_dir) + self.start_index = start_index + self.end_index = end_index + + def test_cuts(self) -> Dict[str, CutSet]: + logging.info("About to get multidataset test cuts") + + dataset_parts = [] + for i in range(self.start_index, self.end_index + 1): + idx = f"{i}".zfill(2) + dataset_parts.append(f"SPEECHIO_ASR_ZH000{idx}") + + prefix="speechio" + suffix="jsonl.gz" + + results_dict = {} + for partition in dataset_parts: + path = f"{prefix}_cuts_{partition}.{suffix}" + + logging.info(f"Loading {path} set in lazy mode") + test_cuts = load_manifest_lazy( + self.fbank_dir / path + ) + results_dict[partition] = test_cuts + + return results_dict \ No newline at end of file diff --git a/egs/speechio/ASR/whisper/requirements.txt b/egs/speechio/ASR/whisper/requirements.txt new file mode 120000 index 0000000000..744bf8bb66 --- /dev/null +++ b/egs/speechio/ASR/whisper/requirements.txt @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/requirements.txt \ No newline at end of file diff --git a/egs/speechio/ASR/whisper/whisper_encoder_forward_monkey_patch.py b/egs/speechio/ASR/whisper/whisper_encoder_forward_monkey_patch.py new file mode 120000 index 0000000000..2a78089212 --- /dev/null +++ b/egs/speechio/ASR/whisper/whisper_encoder_forward_monkey_patch.py @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/whisper_encoder_forward_monkey_patch.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/asr_datamodule.py b/egs/speechio/ASR/zipformer/asr_datamodule.py new file mode 120000 index 0000000000..bf446dabe7 --- /dev/null +++ b/egs/speechio/ASR/zipformer/asr_datamodule.py @@ -0,0 +1 @@ +../whisper/asr_datamodule.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/beam_search.py b/egs/speechio/ASR/zipformer/beam_search.py new file mode 120000 index 0000000000..8e2c0a65c5 --- /dev/null +++ b/egs/speechio/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/beam_search.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/ctc_decode.py b/egs/speechio/ASR/zipformer/ctc_decode.py new file mode 100644 index 0000000000..f9d0db993b --- /dev/null +++ b/egs/speechio/ASR/zipformer/ctc_decode.py @@ -0,0 +1,623 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Liyong Guo, +# Quandong Wang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +(1) ctc-decoding +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method ctc-decoding + +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import AsrDataModule +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import get_lattice, one_best_decoding +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + get_texts, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_2000/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_2000", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="ctc-decoding", + help="""Decoding method. + Supported values are: + - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece + model, i.e., lang_dir/bpe.model, to convert word pieces to words. + It needs neither a lexicon nor an n-gram LM. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for n-best based decoding method. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, and nbest-oracle + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=1.0, + help="""The scale to be applied to `lattice.scores`. + It's needed if you use any kinds of n-best based rescoring. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, and nbest-oracle + A smaller value results in more unique paths. + """, + ) + + add_model_arguments(parser) + + return parser + + +def get_decoding_params() -> AttributeDict: + """Parameters for decoding.""" + params = AttributeDict( + { + "frame_shift_ms": 10, + "search_beam": 20, + "output_beam": 8, + "min_active_states": 30, + "max_active_states": 10000, + "use_double_scores": True, + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + batch: dict, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + - key: It indicates the setting used for decoding. For example, + if no rescoring is used, the key is the string `no_rescore`. + If LM rescoring is used, the key is the string `lm_scale_xxx`, + where `xxx` is the value of `lm_scale`. An example key is + `lm_scale_0.7` + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + + Args: + params: + It's the return value of :func:`get_params`. + + - params.decoding_method is "1best", it uses 1best decoding without LM rescoring. + - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring. + - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring. + - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM + rescoring. + + model: + The neural model. + H: + The ctc topo. Used only when params.decoding_method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.decoding_method is ctc-decoding. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + G: + An LM. It is not None when params.decoding_method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return the decoding result. See above description for the format of + the returned dict. Note: If it decodes to nothing, then return None. + """ + device = H.device + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens) + ctc_output = model.ctc_output(encoder_out) # (N, T, C) + + supervision_segments = torch.stack( + ( + supervisions["sequence_idx"], + torch.div( + supervisions["start_frame"], + params.subsampling_factor, + rounding_mode="floor", + ), + torch.div( + supervisions["num_frames"], + params.subsampling_factor, + rounding_mode="floor", + ), + ), + 1, + ).to(torch.int32) + + assert bpe_model is not None + decoding_graph = H + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=decoding_graph, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.decoding_method == "ctc-decoding": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + # Note: `best_path.aux_labels` contains token IDs, not word IDs + # since we are using H, not HLG here. + # + # token_ids is a lit-of-list of IDs + token_ids = get_texts(best_path) + + # hyps is a list of str, e.g., ['xxx yyy zzz', ...] + hyps = bpe_model.decode(token_ids) + + # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ] + hyps = [s.split() for s in hyps] + key = "ctc-decoding" + return {key: hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + word_table: k2.SymbolTable, + G: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + HLG: + The decoding graph. Used only when params.decoding_method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.decoding_method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.decoding_method is ctc-decoding. + word_table: + It is the word symbol table. + G: + An LM. It is not None when params.decoding_method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return a dict, whose key may be "no-rescore" if no LM rescoring + is used, or it may be "lm_scale_0.7" if LM rescoring is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + H=H, + bpe_model=bpe_model, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = list(ref_text.replace(" ", "")) + hyp_words = list("".join(hyp_words)) + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % 100 == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}-{key}", results) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + + assert params.decoding_method in ("ctc-decoding",) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + logging.info(params) + + lexicon = Lexicon(params.lang_dir) + max_token_id = max(lexicon.tokens) + num_classes = max_token_id + 1 # +1 for the blank + + params.vocab_size = num_classes + # and are defined in local/train_bpe_model.py + params.blank_id = 0 + + HLG = None + H = k2.ctc_topo( + max_token=max_token_id, + modified=True, + device=device, + ) + bpe_model = spm.SentencePieceProcessor() + bpe_model.load(str(params.lang_dir / "bpe.model")) + + G = None + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir, args.start_index, args.end_index) + + test_sets_cuts = multi_dataset.test_cuts() + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}" + ) + return T > 0 + + test_sets = test_sets_cuts.keys() + test_dl = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dl): + logging.info(f"Start decoding test set: {test_set}") + + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + word_table=lexicon.word_table, + G=G, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/speechio/ASR/zipformer/decode.py b/egs/speechio/ASR/zipformer/decode.py new file mode 100644 index 0000000000..91c43d0448 --- /dev/null +++ b/egs/speechio/ASR/zipformer/decode.py @@ -0,0 +1,828 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(6) fast beam search (nbest oracle WER) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(7) fast beam search (with LG) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import AsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_2000/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_2000", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens) + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_LG": + hyp_tokens = fast_beam_search_nbest_LG( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in hyp_tokens: + hyps.append([word_table[i] for i in hyp]) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + texts = [list(str(text).replace(" ", "")) for text in texts] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + hyp_text = "".join(hyp_words) + this_batch.append((cut_id, ref_text, hyp_text)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir, args.start_index, args.end_index) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}" + ) + return T > 0 + + test_sets_cuts = multi_dataset.test_cuts() + + test_sets = test_sets_cuts.keys() + test_dl = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dl): + logging.info(f"Start decoding test set: {test_set}") + + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/speechio/ASR/zipformer/decoder.py b/egs/speechio/ASR/zipformer/decoder.py new file mode 120000 index 0000000000..5a8018680d --- /dev/null +++ b/egs/speechio/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/encoder_interface.py b/egs/speechio/ASR/zipformer/encoder_interface.py new file mode 120000 index 0000000000..c2eaca6712 --- /dev/null +++ b/egs/speechio/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24 b/egs/speechio/ASR/zipformer/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24 new file mode 160000 index 0000000000..3e03a390b0 --- /dev/null +++ b/egs/speechio/ASR/zipformer/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24 @@ -0,0 +1 @@ +Subproject commit 3e03a390b04d3b0dc91c5681e1c51789e9ad5d66 diff --git a/egs/speechio/ASR/zipformer/joiner.py b/egs/speechio/ASR/zipformer/joiner.py new file mode 120000 index 0000000000..5b8a36332e --- /dev/null +++ b/egs/speechio/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/model.py b/egs/speechio/ASR/zipformer/model.py new file mode 120000 index 0000000000..cd7e07d72b --- /dev/null +++ b/egs/speechio/ASR/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/multi_dataset.py b/egs/speechio/ASR/zipformer/multi_dataset.py new file mode 120000 index 0000000000..af164667a4 --- /dev/null +++ b/egs/speechio/ASR/zipformer/multi_dataset.py @@ -0,0 +1 @@ +../whisper/multi_dataset.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/optim.py b/egs/speechio/ASR/zipformer/optim.py new file mode 120000 index 0000000000..5eaa3cffd4 --- /dev/null +++ b/egs/speechio/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/scaling.py b/egs/speechio/ASR/zipformer/scaling.py new file mode 120000 index 0000000000..6f398f431d --- /dev/null +++ b/egs/speechio/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/scaling_converter.py b/egs/speechio/ASR/zipformer/scaling_converter.py new file mode 120000 index 0000000000..b0ecee05e1 --- /dev/null +++ b/egs/speechio/ASR/zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/subsampling.py b/egs/speechio/ASR/zipformer/subsampling.py new file mode 120000 index 0000000000..01ae9002c6 --- /dev/null +++ b/egs/speechio/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/train.py b/egs/speechio/ASR/zipformer/train.py new file mode 100644 index 0000000000..c1bbd2ee83 --- /dev/null +++ b/egs/speechio/ASR/zipformer/train.py @@ -0,0 +1,1385 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Daniel Povey) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +# For non-streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --max-duration 1000 + +# For streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --causal 1 \ + --max-duration 1000 + +It supports training with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from multi_dataset import MultiDataset +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_2000/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir) + + train_cuts = multi_dataset.train_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = sp.encode(c.supervisions[0].text, out_type=str) + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = data_module.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = multi_dataset.dev_cuts() + valid_dl = data_module.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/speechio/ASR/zipformer/zipformer.py b/egs/speechio/ASR/zipformer/zipformer.py new file mode 120000 index 0000000000..23011dda71 --- /dev/null +++ b/egs/speechio/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file