def test_wav_augment_with_executor(self, exec_type):
cut = self.with_cut(sampling_rate=16000, num_samples=16000)
with TemporaryDirectory() as d, \
LilcomFilesWriter(storage_path=d) as storage, \
exec_type(max_workers=4) as ex:
cut_set = CutSet.from_cuts(cut.with_id(str(i)) for i in range(100))
# Just test that it runs and does not hang.
cut_set_feats = cut_set.compute_and_store_features(
extractor=Fbank(),
storage=storage,
augment_fn=SoxEffectTransform(speed(16000)),
executor=ex)
def main():
dataset_parts = ('dev-clean', 'test-clean', 'train-clean-100')
print("Parts we will prepare: ", dataset_parts)
corpus_dirs = [
Path('/export/corpora5/LibriSpeech'),
Path(
'/home/storage04/zhuangweiji/data/open-source-data/librispeech/LibriSpeech'
)
]
corpus_dir = None
for d in corpus_dirs:
if os.path.exists(d):
corpus_dir = d
if corpus_dir is None:
print(
"Please create a place on your system to put the downloaded Librispeech data "
"and add it to `corpus_dirs`")
sys.exit(1)
output_dir = Path('exp/data')
print('Manifest preparation:')
librispeech_manifests = prepare_librispeech(corpus_dir=corpus_dir,
dataset_parts=dataset_parts,
output_dir=output_dir,
num_jobs=num_jobs)
print('Feature extraction:')
with get_executor() as ex: # Initialize the executor only once.
for partition, manifests in librispeech_manifests.items():
if (output_dir / f'cuts_{partition}.json.gz').is_file():
print(f'{partition} already exists - skipping.')
continue
print('Processing', partition)
cut_set = CutSet.from_manifests(
recordings=manifests['recordings'],
supervisions=manifests['supervisions'])
if 'train' in partition:
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=Fbank(),
executor=ex,
storage=LilcomFilesWriter(f'{output_dir}/feats_{partition}'))
librispeech_manifests[partition]['cuts'] = cut_set
cut_set.to_json(output_dir / f'cuts_{partition}.json.gz')
def test_padded_cut_num_frames_and_samples_are_consistent(
sampling_rate, num_samples, padded_duration):
with make_cut(sampling_rate, num_samples) as cut, \
TemporaryDirectory() as dir, \
LilcomFilesWriter(dir) as storage:
cut = cut.compute_and_store_features(extractor=Fbank(),
storage=storage)
cut = cut.pad(padded_duration)
feats = cut.load_features()
samples = cut.load_audio()
assert cut.has_features
assert feats.shape[0] == cut.num_frames
assert feats.shape[1] == cut.num_features
assert cut.has_recording
assert samples.shape[0] == 1
assert samples.shape[1] == cut.num_samples
def test_mixed_cut_num_frames_example_1():
fbank = Fbank()
with make_cut(sampling_rate=16000, num_samples=237920) as cut1, \
make_cut(sampling_rate=16000, num_samples=219600) as cut2, \
TemporaryDirectory() as d, \
LilcomFilesWriter(d) as storage:
# These are two cuts of similar duration, concatenated together with 1 second of silence
# in between, and padded to duration of 31.445.
mixed: MixedCut = (cut1.compute_and_store_features(
fbank, storage).pad(duration=cut1.duration + 1.0).append(
cut2.compute_and_store_features(fbank,
storage)).pad(duration=31.445))
assert mixed.duration == 31.445 # Padded correctly
assert mixed.num_frames == 3145 # Round last 5 up
assert sum(
t.cut.num_frames for t in mixed.tracks
) == 3145 # Since the tracks do not overlap in this example,
# The sum of individual cut num_frames should be equal to the total num_frames
features = mixed.load_features()
assert features.shape[
0] == 3145 # Loaded features num frames matches the meta-data
def test_mixed_cut_num_frames_example_2():
fbank = Fbank()
with make_cut(sampling_rate=16000, num_samples=252879) as cut1, \
make_cut(sampling_rate=16000, num_samples=185280) as cut2, \
make_cut(sampling_rate=16000, num_samples=204161) as cut3, \
TemporaryDirectory() as d, \
LilcomFilesWriter(d) as storage:
# These are two cuts of similar duration, concatenated together with 1 second of silence
# in between, and padded to duration of 31.445.
mixed: MixedCut = (cut1.compute_and_store_features(
fbank, storage).pad(duration=cut1.duration + 1.0).append(
cut2.compute_and_store_features(fbank, storage)))
mixed = (mixed.pad(duration=mixed.duration + 1.0).append(
cut3.compute_and_store_features(fbank, storage)))
assert mixed.duration == 42.145 # Padded correctly
assert mixed.num_frames == 4215 # Round last 5 up
# TODO(pzelasko): This assertion would not pass for now, as we're adding an extra frame during load_features.
# assert sum(t.cut.num_frames for t in mixed.tracks) == 4215 # Since the tracks do not overlap in this example,
# The sum of individual cut num_frames should be equal to the total num_frames
features = mixed.load_features()
assert features.shape[
0] == 4215 # Loaded features num frames matches the meta-data
def main():
corpus_dirs = [Path('/mnt/cfs2/asr/database/AM/aishell')]
corpus_dir = None
for d in corpus_dirs:
if os.path.exists(d):
corpus_dir = d
if corpus_dir is None:
print(
"Please create a place on your system to put the downloaded Aishell data "
"and add it to `corpus_dirs`")
sys.exit(1)
output_dir = Path('exp/data')
print('Manifest preparation:')
aishell_manifests = prepare_aishell(corpus_dir=corpus_dir,
output_dir=output_dir)
print('Feature extraction:')
with get_executor() as ex: # Initialize the executor only once.
for partition, manifests in aishell_manifests.items():
if (output_dir / f'cuts_{partition}.json.gz').is_file():
print(f'{partition} already exists - skipping.')
continue
print('Processing', partition)
cut_set = CutSet.from_manifests(
recordings=manifests['recordings'],
supervisions=manifests['supervisions'])
if 'train' in partition:
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=Fbank(),
executor=ex,
storage=LilcomFilesWriter(f'{output_dir}/feats_{partition}'))
aishell_manifests[partition]['cuts'] = cut_set
cut_set.to_json(output_dir / f'cuts_{partition}.json.gz')
augmenter = WavAugmenter.create_predefined(
'pitch_reverb_tdrop',
sampling_rate=16000) if use_data_augmentation else None
# It seems when spawning multiple Python subprocesses with the same sox handle it raises "std::runtime_error: Couldn't close file"
# The issue seems to happen only in a Jupyter notebook on Mac OS X, hence the work around below.
if use_data_augmentation:
num_jobs = 1
else:
num_jobs = os.cpu_count()
torch.set_num_threads(1)
torch.set_num_interop_threads(1)
num_jobs = 1
for partition, manifests in librispeech_manifests.items():
print(partition)
with LilcomFilesWriter(f'{output_dir}/feats_{partition}'
) as storage, ProcessPoolExecutor(num_jobs) as ex:
cut_set = CutSet.from_manifests(
recordings=manifests['recordings'],
supervisions=manifests['supervisions']).compute_and_store_features(
extractor=Fbank(),
storage=storage,
augmenter=augmenter if 'train' in partition else None,
executor=ex)
librispeech_manifests[partition]['cuts'] = cut_set
cut_set.to_json(output_dir + f'/cuts_{partition}.json.gz')
cuts_train = SpeechRecognitionDataset(
librispeech_manifests['train-clean-100']['cuts'])
cuts_test = SpeechRecognitionDataset(
librispeech_manifests['test-clean']['cuts'])
def _with_features(self, cut: Cut, frame_shift: Seconds) -> Cut:
d = TemporaryDirectory()
self.dirs.append(d)
extractor = Fbank(config=FbankConfig(frame_shift=frame_shift))
with LilcomFilesWriter(d.name) as storage:
return cut.compute_and_store_features(extractor, storage=storage)
def _with_features(self, cut: Cut) -> Cut:
d = TemporaryDirectory()
self.dirs.append(d)
with LilcomFilesWriter(d.name) as storage:
return cut.compute_and_store_features(Fbank(), storage=storage)