Musan implementation for ReazonSpeech (#1988)

egs/reazonspeech/ASR/local/compute_fbank_musan.py

@@ -0,0 +1,154 @@
+#!/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 computes fbank features of the musan dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/manifests.
+"""
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    LilcomChunkyWriter,
+    MonoCut,
+    WhisperFbank,
+    WhisperFbankConfig,
+    combine,
+)
+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 is_cut_long(c: MonoCut) -> bool:
+    return c.duration > 5
+
+
+def compute_fbank_musan(
+    num_mel_bins: int = 80,
+    whisper_fbank: bool = False,
+    output_dir: str = "data/manifests",
+):
+    src_dir = Path("data/manifests")
+    output_dir = Path(output_dir)
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "music",
+        "speech",
+        "noise",
+    )
+    prefix = "musan"
+    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,
+    )
+
+    musan_cuts_path = output_dir / "musan_cuts.jsonl.gz"
+
+    if musan_cuts_path.is_file():
+        logging.info(f"{musan_cuts_path} already exists - skipping")
+        return
+
+    logging.info("Extracting features for Musan")
+
+    if whisper_fbank:
+        extractor = WhisperFbank(
+            WhisperFbankConfig(num_filters=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.
+        # create chunks of Musan with duration 5 - 10 seconds
+        musan_cuts = (
+            CutSet.from_manifests(
+                recordings=combine(part["recordings"] for part in manifests.values())
+            )
+            .cut_into_windows(10.0)
+            .filter(is_cut_long)
+            .compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/musan_feats",
+                num_jobs=num_jobs if ex is None else 80,
+                executor=ex,
+                storage_type=LilcomChunkyWriter,
+            )
+        )
+        musan_cuts.to_file(musan_cuts_path)
+
+
+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(
+        "--output-dir",
+        type=str,
+        default="data/manifests",
+        help="Output directory. Default: data/manifests.",
+    )
+    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_musan(
+        num_mel_bins=args.num_mel_bins,
+        whisper_fbank=args.whisper_fbank,
+        output_dir=args.output_dir,
+    )

egs/reazonspeech/ASR/local/utils/asr_datamodule.py

@@ -180,7 +180,10 @@ class ReazonSpeechAsrDataModule:
         )
 
     def train_dataloaders(
-        self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+        cuts_musan: Optional[CutSet] = None,
     ) -> DataLoader:
         """
         Args:
@@ -191,6 +194,14 @@ class ReazonSpeechAsrDataModule:
         """
 
         transforms = []
+        if cuts_musan is not None:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
+            )
+        else:
+            logging.info("Disable MUSAN")
+
         input_transforms = []
 
         if self.args.enable_spec_aug:

egs/reazonspeech/ASR/prepare.sh

@@ -17,8 +17,16 @@ stop_stage=100
 #      You can find FLAC files in this directory.
 #      You can download them from https://huggingface.co/datasets/reazon-research/reazonspeech
 #
-#  - $dl_dir/dataset.json
+#  - $dl_dir/ReazonSpeech/dataset.json
 #      The metadata of the ReazonSpeech dataset.
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
 
 dl_dir=$PWD/download
 
@@ -48,7 +56,15 @@ if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
   #
   if [ ! -d $dl_dir/ReazonSpeech/downloads ]; then
     # Download small-v1 by default.
-    lhotse download reazonspeech --subset small-v1 $dl_dir
+    lhotse download reazonspeech --subset medium $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
 
@@ -64,7 +80,18 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
 fi
 
 if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
-    log "Stage 2: Compute ReazonSpeech fbank"
+  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_prep.done ]; then
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan_prep.done
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+    log "Stage 3: Compute ReazonSpeech fbank"
     if [ ! -e data/manifests/.reazonspeech-validated.done ]; then
         python local/compute_fbank_reazonspeech.py --manifest-dir data/manifests
         python local/validate_manifest.py --manifest data/manifests/reazonspeech_cuts_train.jsonl.gz
@@ -74,13 +101,22 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
     fi
 fi
 
-if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
-    log "Stage 3: Prepare ReazonSpeech lang_char"
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  mkdir -p data/manifests
+  if [ ! -e data/manifests/.musan_fbank.done ]; then
+    ./local/compute_fbank_musan.py
+    touch data/manifests/.musan_fbank.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+    log "Stage 5: Prepare ReazonSpeech lang_char"
     python local/prepare_lang_char.py data/manifests/reazonspeech_cuts_train.jsonl.gz
 fi
 
-if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
-    log "Stage 4: Show manifest statistics"
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+    log "Stage 6: Show manifest statistics"
     python local/display_manifest_statistics.py --manifest-dir data/manifests > data/manifests/manifest_statistics.txt
     cat data/manifests/manifest_statistics.txt
-fi
+fi

egs/reazonspeech/ASR/zipformer/train.py

@@ -65,6 +65,7 @@ import torch.nn as nn
 from asr_datamodule import ReazonSpeechAsrDataModule
 from decoder import Decoder
 from joiner import Joiner
+from lhotse import load_manifest
 from lhotse.cut import Cut
 from lhotse.dataset.sampling.base import CutSampler
 from lhotse.utils import fix_random_seed
@@ -1219,8 +1220,23 @@ def run(rank, world_size, args):
     else:
         sampler_state_dict = None
 
+    if args.enable_musan:
+        musan_path = Path(args.manifest_dir) / "musan_cuts.jsonl.gz"
+        if musan_path.exists():
+            cuts_musan = load_manifest(musan_path)
+            logging.info(f"Loaded MUSAN manifest from {musan_path}")
+        else:
+            logging.warning(
+                f"MUSAN manifest not found at {musan_path}, disabling MUSAN augmentation"
+            )
+            cuts_musan = None
+    else:
+        cuts_musan = None
+
     train_dl = reazonspeech_corpus.train_dataloaders(
-        train_cuts, sampler_state_dict=sampler_state_dict
+        train_cuts,
+        sampler_state_dict=sampler_state_dict,
+        cuts_musan=cuts_musan,
     )
 
     valid_cuts = reazonspeech_corpus.valid_cuts()