Merge 25f465c8c061a04b74b3a71a24215a691b863c39 into 34fc1fdf0d8ff520e2bb18267d046ca207c78ef9

egs/mls/ASR/local/compute_fbank_mls.py

@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+# Copyright    2024  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# 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 LibriSpeech dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import sentencepiece as spm
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, 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)
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--manifest-dir",
+        type=str,
+        default="data/manifests",
+    )
+
+    parser.add_argument(
+        "--fbank-dir",
+        type=str,
+        default="data/fbank_mls",
+    )
+
+    parser.add_argument(
+        "--part",
+        type=str,
+        help="Which language to prepare, if all, prepare all languages",
+        choices=["english", "dutch", "german", "spanish", "french", "italian", "polish", "portuguese", "all"]
+    )
+
+    return parser.parse_args()
+
+def compute_fbank_mls(
+    manifest_dir=str,
+    fbank_dir=str,
+    part=str,
+):
+    src_dir = Path("data/manifests")
+    output_dir = Path(fbank_dir)
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    if part == "all":
+        dataset_parts = [
+            "english",
+            "dutch",
+            "german",
+            "spanish"
+        ]
+    else:
+        dataset_parts = [part]
+    splits = ["train", "test", "dev"]
+
+    num_jobs = 15
+    num_mel_bins = 80
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    for language in dataset_parts:
+        for split in splits:
+            recording_file = src_dir / f"mls-{language}_recordings_{split}.jsonl.gz"
+            supervision_file = src_dir / f"mls-{language}_supervisions_{split}.jsonl.gz"
+            recordings = CutSet.from_file(recording_file)
+            supervisions = CutSet.from_file(supervision_file)
+
+            cut_set = CutSet.from_manifests(
+                recordings=recordings,
+                supervisions=supervisions,
+            )
+
+            prefix = f"mls-{language}"
+            with get_executor() as ex:
+                cut_set = cut_set.compute_and_store_features(
+                    extractor=extractor,
+                    storage_path=f"{output_dir}/{prefix}_feats_{split}",
+                    # when an executor is specified, make more partitions
+                    num_jobs=num_jobs if ex is None else 80,
+                    executor=ex,
+                    storage_type=LilcomChunkyWriter,
+                )
+
+            cuts_filename = output_dir / f"mls-{language}_{split}.jsonl.gz"
+            
+            logging.info(f"Saving to {cuts_filename}")
+            cut_set.to_file(cuts_filename)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    args = get_args()
+    logging.info(vars(args))
+    compute_fbank_mls(
+        manifest_dir=args.manifest_dir,
+        fbank_dir=args.fbank_dir,
+        part=args.part,
+    )
+
+

egs/mls/ASR/local/compute_fbank_mls_splits.py

@@ -0,0 +1,172 @@
+#!/usr/bin/env python3
+# Copyright    2021  Johns Hopkins University (Piotr Żelasko)
+# Copyright    2021  Xiaomi Corp.             (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.
+
+import argparse
+import logging
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig
+
+# 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)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=20,
+        help="Number of dataloading workers used for reading the audio.",
+    )
+    parser.add_argument(
+        "--batch-duration",
+        type=float,
+        default=600.0,
+        help="The maximum number of audio seconds in a batch."
+        "Determines batch size dynamically.",
+    )
+
+    parser.add_argument(
+        "--language",
+        type=str,
+        default="english",
+    )
+
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the English subset",
+    )
+
+    parser.add_argument(
+        "--start",
+        type=int,
+        default=0,
+        help="Process pieces starting from this number (inclusive).",
+    )
+
+    parser.add_argument(
+        "--stop",
+        type=int,
+        default=-1,
+        help="Stop processing pieces until this number (exclusive).",
+    )
+
+    parser.add_argument(
+        "--fbank-dir",
+        type=str,
+        default="data/fbank_mls"
+    )
+    return parser
+
+
+def compute_fbank_mls_splits(args):
+    num_splits = args.num_splits
+    output_dir = f"{args.fbank_dir}/{args.language}_split"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = 8  # num_digits is fixed by lhotse split-lazy
+
+    start = args.start
+    stop = args.stop
+    if stop < start:
+        stop = num_splits
+
+    stop = min(stop, num_splits)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+    extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device))
+    logging.info(f"device: {device}")
+
+    for i in range(start, stop):
+        idx = f"{i}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"mls-{args.language}_train.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"mls-{args.language}_train_raw.{idx}.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Computing features")
+
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/feats_{args.language}_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            overwrite=True,
+        )
+
+        logging.info("About to split cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+        logging.info(f"Saved to {cuts_path}")
+
+
+def main():
+    now = datetime.now()
+    date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
+
+    log_filename = "log-compute_fbank_mls_splits"
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    log_filename = f"{log_filename}-{date_time}"
+
+    logging.basicConfig(
+        filename=log_filename,
+        format=formatter,
+        level=logging.INFO,
+        filemode="w",
+    )
+
+    console = logging.StreamHandler()
+    console.setLevel(logging.INFO)
+    console.setFormatter(logging.Formatter(formatter))
+    logging.getLogger("").addHandler(console)
+
+    parser = get_parser()
+    args = parser.parse_args()
+    logging.info(vars(args))
+
+    compute_fbank_mls_splits(args)
+
+
+if __name__ == "__main__":
+    main()

egs/mls/ASR/local/train_bpe_model.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# Copyright    2024  Xiaomi Corp.        (authors: Xiaoyu Yang)
+#
+# 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.
+
+
+# You can install sentencepiece via:
+#
+#  pip install sentencepiece
+#
+# Due to an issue reported in
+# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030
+#
+# Please install a version >=0.1.96
+
+import argparse
+import shutil
+from pathlib import Path
+
+import sentencepiece as spm
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        help="""Input and output directory.
+        The generated bpe.model is saved to this directory.
+        """,
+    )
+
+    parser.add_argument(
+        "--byte-fallback",
+        action="store_true",
+        help="""Whether to enable byte_fallback when training bpe.""",
+    )
+
+    parser.add_argument(
+        "--character-coverage",
+        type=float,
+        default=1.0,
+        help="Character coverage in vocabulary.",
+    )
+
+    parser.add_argument(
+        "--transcript",
+        type=str,
+        help="Training transcript.",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        help="Vocabulary size for BPE training",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    vocab_size = args.vocab_size
+    lang_dir = Path(args.lang_dir)
+
+    model_type = "bpe"
+
+    model_prefix = f"{lang_dir}/{model_type}_{vocab_size}"
+    train_text = args.transcript
+    input_sentence_size = 100000000
+
+    user_defined_symbols = ["<blk>", "<sos/eos>"]
+    unk_id = len(user_defined_symbols)
+    # Note: unk_id is fixed to 2.
+    # If you change it, you should also change other
+    # places that are using it.
+
+    model_file = Path(model_prefix + ".model")
+    if not model_file.is_file():
+        spm.SentencePieceTrainer.train(
+            input=train_text,
+            vocab_size=vocab_size,
+            model_type=model_type,
+            model_prefix=model_prefix,
+            input_sentence_size=input_sentence_size,
+            character_coverage=args.character_coverage,
+            user_defined_symbols=user_defined_symbols,
+            byte_fallback=args.byte_fallback,
+            unk_id=unk_id,
+            bos_id=-1,
+            eos_id=-1,
+        )
+    else:
+        print(f"{model_file} exists - skipping")
+        return
+
+    shutil.copyfile(model_file, f"{lang_dir}/bpe.model")
+
+
+if __name__ == "__main__":
+    main()

egs/mls/ASR/local/validate_manifest.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/validate_manifest.py

egs/mls/ASR/prepare.sh

@@ -0,0 +1,227 @@
+#!/usr/bin/env bash
+
+# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
+export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+
+set -eou pipefail
+
+nj=15
+# run step 0 to step 5 by default
+stage=0
+stop_stage=5
+
+# Note: This script just prepare the minimal requirements that needed by a
+# transducer training with bpe units.
+#
+# If you want to use ngram or nnlm, please continue running prepare_lm.sh after
+# you succeed running this script.
+#
+# This script also contains the steps to generate phone based units, but they
+# will not run automatically, you can generate the phone based units by
+# bash prepare.sh --stage -1 --stop-stage -1
+# bash prepare.sh --stage 6 --stop-stage 6
+
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/LibriSpeech
+#      You can find BOOKS.TXT, test-clean, train-clean-360, etc, inside it.
+#      You can download them from https://www.openslr.org/12
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+num_per_split=4000
+fbank_dir=data/fbank_mls
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/lang_bpe_xxx,
+# data/lang_bpe_yyy if the array contains xxx, yyy
+vocab_sizes=(
+  # 5000
+  2000
+  1000
+  # 500
+)
+
+# 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 "Running prepare.sh"
+
+log "dl_dir: $dl_dir"
+log "fbank_dir: $fbank_dir"
+
+languages=(
+  english 
+  german
+  dutch
+  spanish
+  italian
+  french
+  polish
+  portuguese
+)
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/MLS,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/MLS $dl_dir/MLS
+  #
+  if [ ! -d $dl_dir/MLS/train-other-500 ]; then
+    lhotse download mls --full $dl_dir
+  fi
+
+  # If you have pre-downloaded it to /path/to/musan,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/musan $dl_dir/
+  #
+  if [ ! -d $dl_dir/musan ]; then
+    lhotse download musan $dl_dir
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare MLS manifest"
+  # We assume that you have downloaded the MLS corpus
+  # to $dl_dir/MLS
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.mls.done ]; then
+    lhotse prepare mls -j $nj $dl_dir/MLS data/manifests
+    touch data/manifests/.mls.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to $dl_dir/musan
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.musan.done ]; then
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Split english subset into pieces (may take 30 minutes)"
+  split_dir=${fbank_dir}/english_split
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split-lazy ${fbank_dir}/mls-english_train_raw.jsonl.gz $split_dir $num_per_split
+    touch $split_dir/.split_completed
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for MLS (except English)"
+  mkdir -p ${fbank_dir}
+  if [ ! -e ${fbank_dir}/.mls.done ]; then
+    ./local/compute_fbank_mls.py
+    touch ${fbank_dir}/.mls.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+log "Stage 5: Compute fbank for English split of MLS"
+  if [ ! -e ${fbank_dir}/.mls-english.done ]; then
+    num_splits=$(find ${fbank_dir}/english_split -name "mls-english_train_raw.*.jsonl.gz" | wc -l)
+    ./local/compute_fbank_mls_splits.py \
+      --fbank-dir $fbank_dir \
+      --num-workers 20 \
+      --language english \
+      --num-splits $num_splits \
+
+    touch ${fbank_dir}/.mls-english.done
+  fi
+
+  if [ ! -e ${fbank_dir}/mls-english_train.jsonl.gz ]; then
+    pieces=$(find ${fbank_dir}/english_split -name "mls-english_train.*.jsonl.gz")
+    lhotse combine $pieces ${fbank_dir}/mls-english_train.jsonl.gz
+  fi
+
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Validate the manifest of MLS"
+  if [ ! -e ${fbank_dir}/.mls-validated.done ]; then
+    log "Validating the fbank features for MLS"
+    parts=(
+      train
+      dev
+      test
+    )
+    for lan in ${languages[@]}; do
+      for part in ${parts[@]}; do
+        python3 ./local/validate_manifest.py \
+          ${fbank_dir}/mls-${lan}_${part}.jsonl.gz
+      done
+    done
+    touch ${fbank_dir}/.mls-validated.done
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Compute fbank for musan"
+  mkdir -p ${fbank_dir}
+  if [ ! -e ${fbank_dir}/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch ${fbank_dir}/.musan.done
+  fi
+fi
+
+if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
+  log "Stage 8: Prepare BPE based lang"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    lang_dir=data/lang_bpe_${vocab_size}
+    mkdir -p $lang_dir
+
+    if [ ! -f $lang_dir/transcript_words.txt ]; then
+      log "Generate data for BPE training"
+      files=(
+        "$dl_dir/MLS/mls_english/train/transcripts.txt"
+        "$dl_dir/MLS/mls_german/train/transcripts.txt"
+        "$dl_dir/MLS/mls_dutch/train/transcripts.txt"
+        "$dl_dir/MLS/mls_french/train/transcripts.txt"
+        "$dl_dir/MLS/mls_spanish/train/transcripts.txt"
+        "$dl_dir/MLS/mls_italian/train/transcripts.txt"
+        "$dl_dir/MLS/mls_portuguese/train/transcripts.txt"
+        "$dl_dir/MLS/mls_polish/train/transcripts.txt"
+      )
+      for f in ${files[@]}; do
+        head -n 1000000 $f | cut -d " " -f 2-
+      done > $lang_dir/transcript_words.txt
+    fi
+
+    if [ ! -f $lang_dir/bpe.model ]; then
+      ./local/train_bpe_model.py \
+        --lang-dir $lang_dir \
+        --vocab-size $vocab_size \
+        --character-coverage 0.999 \
+        --transcript $lang_dir/transcript_words.txt \
+        --byte-fallback 
+    fi
+  done
+fi
+

egs/mls/ASR/shared

@@ -0,0 +1 @@
+../../../icefall/shared

egs/mls/ASR/zipformer/asr_datamodule.py

@@ -0,0 +1,441 @@
+# Copyright      2021  Piotr Żelasko
+# Copyright      2024  Xiaomi Corporation     (Author: Xiaoyu Yang)
+#
+# 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, Fbank, FbankConfig, 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,
+    OnTheFlyFeatures,
+)
+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 MLSAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+
+    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(
+            "--language",
+            type=str,
+            default="all",
+            # choices=["english", "german", "dutch", "french", "spanish", "italian", "portuguese", "polish", "all"],
+            help="""A list of languages separated by comma. If all, use all
+            the languages""",
+        )
+        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=200.0,
+            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(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        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(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            logging.info("About to get Musan cuts")
+            cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                buffer_size=self.args.num_buckets * 2000,
+                shuffle_buffer_size=self.args.num_buckets * 5000,
+                drop_last=self.args.drop_last,
+            )
+        else:
+            logging.info("Using SimpleCutSampler.")
+            train_sampler = SimpleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_sampler.load_state_dict(sampler_state_dict)
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+        valid_dl = DataLoader(
+            validate,
+            sampler=valid_sampler,
+            batch_size=None,
+            num_workers=2,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else 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
+
+    @lru_cache()
+    def train_mls_cuts(self) -> CutSet:
+        if self.args.language == "all":
+            languages = ["english", "german", "dutch", "french", "spanish", "italian", "portuguese", "polish",]
+        else:
+            languages = self.args.language.split(",")
+        if len(languages) == 1:
+            l = languages[0]
+            logging.info(f"About to get {l} cuts")
+            return load_manifest_lazy(
+                self.args.manifest_dir / f"mls-{l}_train.jsonl.gz"
+            )
+        else:
+            all_cuts = []
+            all_cuts_len = []
+            for l in languages:
+                logging.info(f"About to get {l} cuts")
+                current_cuts = load_manifest_lazy(
+                    self.args.manifest_dir / f"mls-{l}_train.jsonl.gz"
+                )
+                current_cuts_len = len(current_cuts)
+                all_cuts.append(current_cuts)
+                all_cuts_len.append(current_cuts_len)
+            return CutSet.mux(
+                *all_cuts,
+                weights=all_cuts_len,
+                stop_early=True,
+            )
+
+    @lru_cache()
+    def mls_dev_cuts(self, language: str) -> CutSet:
+        logging.info(f"About to get dev cuts for {language}")
+        return load_manifest_lazy(
+            self.args.manifest_dir / f"mls-{language}_dev.jsonl.gz"
+        )
+
+    @lru_cache()
+    def mls_test_cuts(self, language: str) -> CutSet:
+        logging.info(f"About to get test cuts for {language}")
+        return load_manifest_lazy(
+            self.args.manifest_dir / f"mls-{language}_test.jsonl.gz"
+        )

egs/mls/ASR/zipformer/beam_search.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py

egs/mls/ASR/zipformer/decoder.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/decoder.py

egs/mls/ASR/zipformer/encoder_interface.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py

egs/mls/ASR/zipformer/joiner.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/joiner.py

egs/mls/ASR/zipformer/model.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/model.py

egs/mls/ASR/zipformer/optim.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/optim.py

egs/mls/ASR/zipformer/scaling.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling.py

egs/mls/ASR/zipformer/scaling_converter.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/scaling_converter.py

egs/mls/ASR/zipformer/subsampling.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/subsampling.py

egs/mls/ASR/zipformer/zipformer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/zipformer/zipformer.py