mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 10:02:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge c934cc2ac448e8b0521aeea1ab0dbfba397c961d into fba5e67d5e14c808cea7f2bf5ccc7fa0c248cc5c

Browse Source
This commit is contained in:
zr_jin 2025-07-05 02:50:54 +08:00 committed by GitHub
commit 9a5197b560
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
9 changed files with 533 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
egs/swbd/ASR/README.md
View File

@ -8,9 +8,9 @@ Switchboard is a collection of about 2,400 two-sided telephone conversations amo
## Performance Record
| | eval2000 | rt03 |
|--------------------------------|------------|--------|
| `conformer_ctc` | 33.37 | 35.06 |
| | eval2000-swbd | eval2000-callhome | eval2000-avg |
|--------------------------------|-----------------|---------------------|--------------|
| `conformer_ctc` | 9.48 | 17.73 | 13.67 |
See [RESULTS](/egs/swbd/ASR/RESULTS.md) for details.

14
egs/swbd/ASR/RESULTS.md
View File

@ -1,6 +1,19 @@
## Results
### Switchboard BPE training results (Conformer-CTC)
#### 2023-12-05 (Narrowband Setup)
The best WER, for the narrowband Switchboard system is presented below
Results using attention decoder are given as:
| | eval2000-swbd | eval2000-callhome | eval2000-avg |
|--------------------------------|-----------------|---------------------|--------------|
| `conformer_ctc` | 11.82 | 23.34 | 17.61 |
Decoding results and models can be found here:
https://huggingface.co/zrjin/icefall-asr-swbd-narrowband-conformer-ctc-2023-12-3
#### 2023-09-04
The best WER, as of 2023-09-04, for the Switchboard is below
@ -13,6 +26,7 @@ Results using attention decoder are given as:
Decoding results and models can be found here:
https://huggingface.co/zrjin/icefall-asr-swbd-conformer-ctc-2023-8-26
#### 2023-06-27
The best WER, as of 2023-06-27, for the Switchboard is below

139
egs/swbd/ASR/local/compute_fbank_eval2000_nb.py Executable file
View File

@ -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,
)

1
egs/swbd/ASR/local/compute_fbank_musan.py Symbolic link
View File

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

110
egs/swbd/ASR/local/compute_fbank_musan_nb.py Executable file
View File

@ -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()

2
egs/swbd/ASR/local/compute_fbank_swbd.py
View File

@ -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.""",
)

162
egs/swbd/ASR/local/compute_fbank_swbd_nb.py Executable file
View File

@ -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,
)

7
egs/swbd/ASR/prepare.sh
View File

@ -145,6 +145,13 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
fi
fi
log "
Computing fbank for SwitchBoard and MUSAN noise.
Note that the current setup upsamples the audio to 16kHz before fbank extraction
please use prepare_nb.sh if you want to use 8kHz audio for narrowband systems.
"
if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
log "Stage 3 I: Compute fbank for SwitchBoard"
if [ ! -e data/fbank/.swbd.done ]; then

96
egs/swbd/ASR/prepare_nb.sh Executable file
View File

@ -0,0 +1,96 @@
#!/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
stage=-1
stop_stage=100
# We assume dl_dir (download dir) contains the following
# directories and files. Most of them can't be downloaded automatically
# as they are not publically available and require a license purchased
# from the LDC.
#
# - $dl_dir/musan
# This directory contains the following directories downloaded from
# http://www.openslr.org/17/
#
# - music
# - noise
# - speech
dl_dir=./download
# swbd1_dir="/export/corpora3/LDC/LDC97S62"
swbd1_dir=./download/LDC97S62/
# eval2000_dir contains the following files and directories
# downloaded from LDC website:
# - LDC2002S09
# - hub5e_00
# - LDC2002T43
# - reference
eval2000_dir="/export/corpora2/LDC/eval2000"
rt03_dir="/export/corpora/LDC/LDC2007S10"
fisher_dir="/export/corpora3/LDC/LDC2004T19"
. shared/parse_options.sh || exit 1
# vocab size for sentence piece models.
# It will generate data/lang_bpe_xxx,
# data/lang_bpe_yyy if the array contains xxx, yyy
vocab_sizes=(
# 5000
# 2000
1000
500
)
# All files generated by this script are saved in "data".
# You can safely remove "data" and rerun this script to regenerate it.
mkdir -p data
log() {
# This function is from espnet
local fname=${BASH_SOURCE[1]##*/}
echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
}
log "swbd1_dir: $swbd1_dir"
log "eval2000_dir: $eval2000_dir"
log "rt03_dir: $rt03_dir"
if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
log "Stage 1 I: Compute narrowband fbank for SwitchBoard"
if [ ! -e data/fbank_nb/.swbd.done ]; then
mkdir -p data/fbank_nb/swbd_split${num_splits}/
for index in $(seq 1 16); do
./local/compute_fbank_swbd_nb.py --split-index ${index} &
done
wait
pieces=$(find data/fbank_nb/swbd_split${num_splits} -name "swbd_cuts_all.*.jsonl.gz")
lhotse combine $pieces data/fbank_nb/swbd_cuts_all.jsonl.gz
touch data/fbank_nb/.swbd.done
fi
fi
if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
log "Stage 1 II: Compute narrowband fbank for eval2000"
if [ ! -e data/fbank_nb/.eval2000.done ]; then
mkdir -p data/fbank_nb/eval2000/
./local/compute_fbank_eval2000_nb.py
touch data/fbank_nb/.eval2000.done
fi
fi
if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
log "Stage 2: Compute narrowband fbank for musan"
mkdir -p data/fbank_nb/
if [ ! -e data/fbank_nb/.musan.done ]; then
./local/compute_fbank_musan_nb.py
touch data/fbank_nb/.musan.done
fi
fi