Add Common Voice (#994)

* Add commonvoice

* Add data preparation recipe

* Updata

* update prepare.sh

* Fix for black

* Update prefix with cv-

* 20 ->

* Update compute_fbank_commonvoice_dev_test.py

* Update prepare.sh

* Update compute_fbank_commonvoice_dev_test.py

egs/commonvoice/ASR/local/compute_fbank_commonvoice_dev_test.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Yifan 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 CommonVoice dataset.
+It looks for manifests in the directory data/${lang}/manifests.
+
+The generated fbank features are saved in data/${lang}/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+from typing import Optional
+
+import torch
+from filter_cuts import filter_cuts
+from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter
+
+# 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(
+        "--language",
+        type=str,
+        help="""Language of Common Voice""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_commonvoice_dev_test(language: str):
+    src_dir = Path(f"data/{language}/manifests")
+    output_dir = Path(f"data/{language}/fbank")
+    num_workers = 42
+    batch_duration = 600
+
+    subsets = ("dev", "test")
+
+    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 partition in subsets:
+        cuts_path = output_dir / f"cv-{language}_cuts_{partition}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping.")
+            continue
+
+        raw_cuts_path = output_dir / f"cv-{language}_cuts_{partition}_raw.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/cv-{language}_feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+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_commonvoice_dev_test(language=args.language)

egs/commonvoice/ASR/local/compute_fbank_commonvoice_splits.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.             (Yifan 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 logging
+from datetime import datetime
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    KaldifeatFbank,
+    KaldifeatFbankConfig,
+    LilcomChunkyWriter,
+    set_audio_duration_mismatch_tolerance,
+    set_caching_enabled,
+)
+
+# 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(
+        "--language",
+        type=str,
+        help="""Language of Common Voice""",
+    )
+
+    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(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the train 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).",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_commonvoice_splits(args):
+    subset = "train"
+    num_splits = args.num_splits
+    language = args.language
+    output_dir = f"data/{language}/fbank/{subset}_split_{num_splits}"
+    output_dir = Path(output_dir)
+    assert output_dir.exists(), f"{output_dir} does not exist!"
+
+    num_digits = len(str(num_splits))
+
+    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}")
+
+    set_audio_duration_mismatch_tolerance(0.01)  # 10ms tolerance
+    set_caching_enabled(False)
+    for i in range(start, stop):
+        idx = f"{i + 1}".zfill(num_digits)
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"cv-{language}_cuts_{subset}.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"cv-{language}_cuts_{subset}_raw.{idx}.jsonl.gz"
+
+        logging.info(f"Loading {raw_cuts_path}")
+        cut_set = CutSet.from_file(raw_cuts_path)
+
+        logging.info("Splitting cuts into smaller chunks.")
+        cut_set = cut_set.trim_to_supervisions(
+            keep_overlapping=False, min_duration=None
+        )
+
+        logging.info("Computing features")
+        cut_set = cut_set.compute_and_store_features_batch(
+            extractor=extractor,
+            storage_path=f"{output_dir}/cv-{language}_feats_{subset}_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+            storage_type=LilcomChunkyWriter,
+            overwrite=True,
+        )
+
+        logging.info(f"Saving to {cuts_path}")
+        cut_set.to_file(cuts_path)
+
+
+def 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_commonvoice_splits(args)
+
+
+if __name__ == "__main__":
+    main()

egs/commonvoice/ASR/local/compute_fbank_musan.py

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

egs/commonvoice/ASR/local/filter_cuts.py

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

egs/commonvoice/ASR/local/preprocess_commonvoice.py

@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+# Copyright    2023  Xiaomi Corp.        (authors: Yifan 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 logging
+from pathlib import Path
+from typing import Optional
+
+from lhotse import CutSet, SupervisionSegment
+from lhotse.recipes.utils import read_manifests_if_cached
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--language",
+        type=str,
+        help="""Language of Common Voice""",
+    )
+
+    return parser.parse_args()
+
+
+def preprocess_commonvoice(
+    language: str,
+    dataset: Optional[str] = None,
+):
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank")
+    output_dir.mkdir(exist_ok=True)
+
+    if dataset is None:
+        dataset_parts = (
+            "dev",
+            "test",
+            "train",
+        )
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    logging.info("Loading manifest")
+    prefix = f"cv-{language}"
+    suffix = "jsonl.gz"
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix=suffix,
+        prefix=prefix,
+    )
+    assert manifests is not None
+
+    assert len(manifests) == len(dataset_parts), (
+        len(manifests),
+        len(dataset_parts),
+        list(manifests.keys()),
+        dataset_parts,
+    )
+
+    for partition, m in manifests.items():
+        logging.info(f"Processing {partition}")
+        raw_cuts_path = output_dir / f"{prefix}_cuts_{partition}_raw.{suffix}"
+        if raw_cuts_path.is_file():
+            logging.info(f"{partition} already exists - skipping")
+            continue
+
+        # Create long-recording cut manifests.
+        cut_set = CutSet.from_manifests(
+            recordings=m["recordings"],
+            supervisions=m["supervisions"],
+        ).resample(16000)
+
+        # Run data augmentation that needs to be done in the
+        # time domain.
+        if "train" in partition:
+            logging.info(
+                f"Speed perturb for {partition} with factors 0.9 and 1.1 "
+                "(Perturbing may take 2 minutes and saving may take 7 minutes)"
+            )
+            cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+        logging.info(f"Saving to {raw_cuts_path}")
+        cut_set.to_file(raw_cuts_path)
+
+
+def 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))
+    preprocess_commonvoice(
+        language=args.language,
+        dataset=args.dataset,
+    )
+    logging.info("Done")
+
+
+if __name__ == "__main__":
+    main()

egs/commonvoice/ASR/prepare.sh

@@ -0,0 +1,156 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+nj=16
+stage=-1
+stop_stage=100
+
+# Split data/${lang}set to this number of pieces
+# This is to avoid OOM during feature extraction.
+num_splits=1000
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/$release/$lang
+#      This directory contains the following files downloaded from
+#       https://mozilla-common-voice-datasets.s3.dualstack.us-west-2.amazonaws.com/${release}/${release}-${lang}.tar.gz
+#
+#     - clips
+#     - dev.tsv
+#     - invalidated.tsv
+#     - other.tsv
+#     - reported.tsv
+#     - test.tsv
+#     - train.tsv
+#     - validated.tsv
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+release=cv-corpus-13.0-2023-03-09
+lang=en
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/${lang}/lang_bpe_xxx,
+# data/${lang}/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/${lang}".
+# You can safely remove "data/${lang}" and rerun this script to regenerate it.
+mkdir -p data/${lang}
+
+log() {
+  # This function is from espnet
+  local fname=${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 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  # If you have pre-downloaded it to /path/to/$release,
+  # you can create a symlink
+  #
+  #   ln -sfv /path/to/$release $dl_dir/$release
+  #
+  if [ ! -d $dl_dir/$release/$lang/clips ]; then
+    lhotse download commonvoice --languages $lang --release $release $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 CommonVoice manifest"
+  # We assume that you have downloaded the CommonVoice corpus
+  # to $dl_dir/$release
+  mkdir -p data/${lang}/manifests
+  if [ ! -e data/${lang}/manifests/.cv-${lang}.done ]; then
+    lhotse prepare commonvoice --language $lang -j $nj $dl_dir/$release data/${lang}/manifests
+    touch data/${lang}/manifests/.cv-${lang}.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 data/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 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Preprocess CommonVoice manifest"
+  if [ ! -e data/${lang}/fbank/.preprocess_complete ]; then
+    ./local/preprocess_commonvoice.py  --language $lang
+    touch data/${lang}/fbank/.preprocess_complete
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for dev and test subsets of CommonVoice"
+  mkdir -p data/${lang}/fbank
+  if [ ! -e data/${lang}/fbank/.cv-${lang}_dev_test.done ]; then
+    ./local/compute_fbank_commonvoice_dev_test.py --language $lang
+    touch data/${lang}/fbank/.cv-${lang}_dev_test.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Split train subset into ${num_splits} pieces"
+  split_dir=data/${lang}/fbank/train_split_${num_splits}
+  if [ ! -e $split_dir/.cv-${lang}_train_split.done ]; then
+    lhotse split $num_splits ./data/${lang}/fbank/cv-${lang}_cuts_train_raw.jsonl.gz $split_dir
+    touch $split_dir/.cv-${lang}_train_split.done
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Compute features for train subset of CommonVoice"
+  if [ ! -e data/${lang}/fbank/.cv-${lang}_train.done ]; then
+    ./local/compute_fbank_commonvoice_splits.py \
+      --num-workers $nj \
+      --batch-duration 600 \
+      --start 0 \
+      --num-splits $num_splits \
+      --language $lang
+    touch data/${lang}/fbank/.cv-${lang}_train.done
+  fi
+fi
+
+if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
+  log "Stage 7: Compute fbank for musan"
+  mkdir -p data/fbank
+  if [ ! -e data/fbank/.musan.done ]; then
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.musan.done
+  fi
+fi

egs/commonvoice/ASR/shared

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