init commit

egs/swbd/ASR/local/compute_fbank_eval2000_nb.py

@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# Modified     2023  The Chinese University of Hong Kong (author: 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 SwitchBoard 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(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=False,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_switchboard(
+    dir_name: str,
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path(f"data/manifests/{dir_name}")
+    output_dir = Path(f"data/fbank_nb/{dir_name}")
+    num_jobs = min(1, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = ("all",)
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = dir_name
+    suffix = "jsonl.gz"
+    manifests = {
+        "eval2000": "data/manifests/eval2000/eval2000_cuts_all_trimmed.jsonl.gz",
+    }
+    assert manifests is not None
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=8000))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        partition = "all"
+        cuts_filename = f"{prefix}_cuts_{partition}.{suffix}"
+        print(cuts_filename)
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{prefix} already exists - skipping.")
+            return
+        logging.info(f"Processing {prefix}")
+        cut_set = CutSet.from_file(manifests[prefix])
+
+        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 = cut_set.trim_to_supervisions(keep_overlapping=False)
+        cut_set.to_file(output_dir / 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_switchboard(
+        dir_name="eval2000",
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )

egs/swbd/ASR/local/compute_fbank_musan.py

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

egs/swbd/ASR/local/compute_fbank_musan_nb.py

@@ -0,0 +1,110 @@
+#!/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/fbank.
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter, MonoCut, combine
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# 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():
+    src_dir = Path("data/manifests")
+    output_dir = Path("data/fbank_nb")
+    num_jobs = min(15, os.cpu_count())
+    num_mel_bins = 80
+
+    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")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=8000))
+
+    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())
+            )
+            .resample(8000)
+            .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)
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    compute_fbank_musan()

egs/swbd/ASR/local/compute_fbank_swbd.py

@@ -66,7 +66,7 @@ def get_args():
     parser.add_argument(
         "--perturb-speed",
         type=str2bool,
-        default=False,
+        default=True,
         help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
     )
 

egs/swbd/ASR/local/compute_fbank_swbd_nb.py

@@ -0,0 +1,162 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# Modified     2023  The Chinese University of Hong Kong (author: 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 SwitchBoard 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(
+        "--bpe-model",
+        type=str,
+        help="""Path to the bpe.model. If not None, we will remove short and
+        long utterances before extracting features""",
+    )
+
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        help="""Dataset parts to compute fbank. If None, we will use all""",
+    )
+
+    parser.add_argument(
+        "--perturb-speed",
+        type=str2bool,
+        default=True,
+        help="""Perturb speed with factor 0.9 and 1.1 on train subset.""",
+    )
+
+    parser.add_argument(
+        "--split-index",
+        type=int,
+        required=True,
+    )
+
+    return parser.parse_args()
+
+
+def compute_fbank_switchboard(
+    dir_name: str,
+    split_index: int,
+    bpe_model: Optional[str] = None,
+    dataset: Optional[str] = None,
+    perturb_speed: Optional[bool] = True,
+):
+    src_dir = Path(f"data/manifests/{dir_name}")
+    output_dir = Path(f"data/fbank_nb/{dir_name}_split16")
+    num_jobs = min(1, os.cpu_count())
+    num_mel_bins = 80
+
+    if bpe_model:
+        logging.info(f"Loading {bpe_model}")
+        sp = spm.SentencePieceProcessor()
+        sp.load(bpe_model)
+
+    if dataset is None:
+        dataset_parts = ("all",)
+    else:
+        dataset_parts = dataset.split(" ", -1)
+
+    prefix = dir_name
+    suffix = "jsonl.gz"
+    split_dir = Path("data/manifests/swbd_split16/")
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=8000))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        partition = "all"
+        cuts_filename = (
+            f"{prefix}_cuts_{partition}.{str(split_index).zfill(2)}.{suffix}"
+        )
+        print(cuts_filename)
+        if (output_dir / cuts_filename).is_file():
+            logging.info(f"{prefix} already exists - skipping.")
+            return
+        logging.info(f"Processing {prefix}")
+        cut_set = (
+            CutSet.from_file(
+                split_dir
+                / f"swbd_train_all_trimmed.{str(split_index).zfill(2)}.jsonl.gz"
+            )
+            .to_eager()
+            .filter(lambda c: c.duration > 2.0)
+        )
+
+        if bpe_model:
+            cut_set = filter_cuts(cut_set, sp)
+        if perturb_speed:
+            logging.info(f"Doing speed perturb")
+            cut_set = cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)
+        cut_set = cut_set.compute_and_store_features(
+            extractor=extractor,
+            storage_path=f"{output_dir}/{prefix}_feats_{partition}_{str(split_index).zfill(2)}",
+            # 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 = cut_set.trim_to_supervisions(
+            keep_overlapping=False,
+            min_duration=None,
+        )
+        cut_set.to_file(output_dir / 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_switchboard(
+        dir_name="swbd",
+        split_index=args.split_index,
+        bpe_model=args.bpe_model,
+        dataset=args.dataset,
+        perturb_speed=args.perturb_speed,
+    )