Feature extraction code for GigaSpeech.

egs/librispeech/ASR/.gitignore

@@ -0,0 +1 @@
+log-*

egs/librispeech/ASR/local/compute_fbank_gigaspeech_dev_test.py

@@ -0,0 +1,92 @@
+#!/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 logging
+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 compute_fbank_gigaspeech_dev_test():
+    in_out_dir = Path("data/fbank")
+    # number of workers in dataloader
+    num_workers = 20
+
+    # number of seconds in a batch
+    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 = in_out_dir / f"cuts_{partition}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = in_out_dir / f"cuts_{partition}_raw.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"{in_out_dir}/feats_{partition}",
+            num_workers=num_workers,
+            batch_duration=batch_duration,
+        )
+        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():
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    compute_fbank_gigaspeech_dev_test()
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/local/compute_fbank_gigaspeech_splits.py

@@ -0,0 +1,168 @@
+#!/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(
+        "--num-splits",
+        type=int,
+        required=True,
+        help="The number of splits of the XL 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
+
+
+def compute_fbank_gigaspeech_splits(args):
+    num_splits = args.num_splits
+    output_dir = f"data/fbank/XL_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}")
+
+    for i in range(start, stop):
+        idx = i
+        logging.info(f"Processing {idx}/{num_splits}")
+
+        cuts_path = output_dir / f"cuts_XL.{idx}.jsonl.gz"
+        if cuts_path.is_file():
+            logging.info(f"{cuts_path} exists - skipping")
+            continue
+
+        raw_cuts_path = output_dir / f"cuts_XL_raw.{idx}.jsonl.gz"
+        if not raw_cuts_path.is_file():
+            logging.info(f"{raw_cuts_path} does not exist - skipping it")
+            continue
+
+        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_XL_{idx}",
+            num_workers=args.num_workers,
+            batch_duration=args.batch_duration,
+        )
+
+        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_gigaspeech_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_gigaspeech_splits(args)
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/local/preprocess_gigaspeech.py

@@ -101,11 +101,6 @@ def preprocess_giga_speech():
                 + cut_set.perturb_speed(0.9)
                 + cut_set.perturb_speed(1.1)
             )
-
-        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 {raw_cuts_path}")
         cut_set.to_file(raw_cuts_path)
 

egs/librispeech/ASR/prepare_giga_speech.sh

@@ -24,6 +24,15 @@ stop_stage=100
 # DEV 12 hours
 # Test 40 hours
 
+# Split XL subset to this number of pieces
+# This is to avoid OOM during feature extraction.
+num_splits=2000
+# We use lazy split from lhotse.
+# The XL subset contains 113916 cuts after speed perturbing with factors
+# 0.9 and 1.1. We want to split it into 2000 splits, so each split
+# contains about 113916 / 2000 = 57 cuts. As a result, there will be 1999 splits.
+chunk_size=57 # number of cuts in each split. The last split may contain fewer cuts.
+
 dl_dir=$PWD/download
 
 . shared/parse_options.sh || exit 1
@@ -107,3 +116,34 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
    touch data/fbank/.preprocess_complete
   fi
 fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Compute features for DEV and TEST subsets of GigaSpeech (may take 2 minutes)"
+  python3 ./local/compute_fbank_gigaspeech_dev_test.py
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Split XL subset into ${num_splits} pieces"
+  split_dir=data/fbank/XL_split_${num_splits}
+  if [ ! -f $split_dir/.split_completed ]; then
+    lhotse split-lazy ./data/fbank/cuts_XL_raw.jsonl.gz $split_dir $chunk_size
+    touch $split_dir/.split_completed
+  fi
+fi
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute features for XL"
+  # Note: The script supports --start and --stop options.
+  # You can use several machines to compute the features in parallel.
+  python3 ./local/compute_fbank_gigaspeech_splits.py \
+    --num-workers $nj \
+    --batch-duration 600 \
+    --num-splits $num_splits
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Combine features for XL"
+  if [ ! -f data/fbank/cuts_XL.jsonl.gz ]; then
+    pieces=$(find data/fbank/XL_split_${num_splits} -name "cuts_XL.*.jsonl.gz")
+    lhotse combine $pieces data/fbank/cuts_XL.jsonl.gz
+  fi
+fi