def test_cut_features_mask(self, supervisions):
cut = Cut('cut', start=0, duration=2, channel=0,
features=Mock(sampling_rate=16000, frame_shift=0.01, num_frames=2000),
supervisions=supervisions)
mask = cut.supervisions_feature_mask()
assert (mask[:50] == 1).all()
assert (mask[50:] == 0).all()
def test_cut_speakers_audio_mask(self, supervisions, alignment):
cut = Cut('cut',
start=0,
duration=2,
channel=0,
recording=Mock(sampling_rate=16000),
supervisions=supervisions)
mask = cut.speakers_audio_mask(use_alignment_if_exists=alignment)
if alignment == "word":
ones = [
np.index_exp[list(chain(range(0, 1600), range(3200, 6400)))],
np.index_exp[list(chain(range(9600, 12800)))]
]
zeros = [
np.index_exp[list(chain(range(1600, 3200), range(6400,
32000)))],
np.index_exp[list(chain(range(0, 9600), range(12800, 32000)))]
]
else:
ones = [
np.index_exp[range(0, 8000)], np.index_exp[range(9600, 12800)]
]
zeros = [
np.index_exp[list(chain(range(8000, 32000)))],
np.index_exp[list(chain(range(0, 9600), range(12800, 32000)))]
]
assert (mask[0, ones[0]] == 1).all()
assert (mask[1, ones[1]] == 1).all()
assert (mask[0, zeros[0]] == 0).all()
assert (mask[1, zeros[1]] == 0).all()
def test_cut_speakers_features_mask(self, supervisions, alignment):
cut = Cut('cut',
start=0,
duration=2,
channel=0,
features=Mock(sampling_rate=16000,
frame_shift=0.01,
num_frames=2000),
supervisions=supervisions)
mask = cut.speakers_feature_mask(use_alignment_if_exists=alignment)
if alignment == "word":
ones = [
np.index_exp[list(chain(range(0, 10), range(20, 40)))],
np.index_exp[list(chain(range(60, 80)))]
]
zeros = [
np.index_exp[list(chain(range(10, 20), range(40, 200)))],
np.index_exp[list(chain(range(0, 60), range(80, 200)))]
]
else:
ones = [
np.index_exp[list(chain(range(0, 50)))],
np.index_exp[list(chain(range(60, 80)))]
]
zeros = [
np.index_exp[list(chain(range(50, 200)))],
np.index_exp[list(chain(range(0, 60), range(80, 200)))]
]
assert (mask[0, ones[0]] == 1).all()
assert (mask[1, ones[1]] == 1).all()
assert (mask[0, zeros[0]] == 0).all()
assert (mask[1, zeros[1]] == 0).all()
def test_cut_audio_mask(self):
cut = Cut('cut',
start=0,
duration=2,
channel=0,
recording=Mock(sampling_rate=16000))
mask = cut.supervisions_audio_mask()
assert mask.sum() == 0
def test_cut_features_mask(self):
cut = Cut('cut',
start=0,
duration=2,
channel=0,
features=Mock(sampling_rate=16000, frame_shift=0.01))
mask = cut.supervisions_feature_mask()
assert mask.sum() == 0
def test_mixed_cut_audio_mask(self, supervisions):
cut = Cut('cut', start=0, duration=2, channel=0, recording=Mock(sampling_rate=16000),
supervisions=supervisions)
mixed_cut = cut.append(cut)
mask = mixed_cut.supervisions_audio_mask()
assert (mask[:8000] == 1).all()
assert (mask[8000:32000] == 0).all()
assert (mask[32000:40000] == 1).all()
assert (mask[40000:] == 0).all()
def test_mixed_cut_features_mask(self, supervisions):
cut = Cut('cut', start=0, duration=2, channel=0, features=Mock(sampling_rate=16000, frame_shift=0.01),
supervisions=supervisions)
mixed_cut = cut.append(cut)
mask = mixed_cut.supervisions_feature_mask()
assert (mask[:50] == 1).all()
assert (mask[50:200] == 0).all()
assert (mask[200:250] == 1).all()
assert (mask[250:] == 0).all()
def with_cut(
self,
sampling_rate: int,
num_samples: int,
features: bool = True,
supervision: bool = False
) -> Cut:
duration = num_samples / sampling_rate
cut = Cut(
id=str(uuid4()),
start=0,
duration=duration,
channel=0,
recording=self.with_recording(sampling_rate=sampling_rate, num_samples=num_samples)
)
if features:
cut = self._with_features(cut)
if supervision:
cut.supervisions.append(SupervisionSegment(
id=f'sup-{cut.id}',
recording_id=cut.recording_id,
start=0,
duration=cut.duration,
text='irrelevant'
))
return cut
def test_mixed_cut_audio_mask(self, supervisions):
cut = Cut('cut',
start=0,
duration=2,
channel=0,
recording=Mock(sampling_rate=16000),
supervisions=supervisions)
mixed_cut = cut.append(cut)
mask = mixed_cut.supervisions_audio_mask()
ones = np.index_exp[list(
chain(range(0, 8000), range(9600, 12800), range(32000, 40000),
range(41600, 44800)))]
zeros = np.index_exp[list(
chain(range(8000, 9600), range(12800, 32000), range(40000, 41600),
range(44800, 64000)))]
assert (mask[ones] == 1).all()
assert (mask[zeros] == 0).all()
def test_mixed_cut_features_mask(self, supervisions):
cut = Cut('cut',
start=0,
duration=2,
channel=0,
features=Mock(sampling_rate=16000, frame_shift=0.01),
supervisions=supervisions)
mixed_cut = cut.append(cut)
mask = mixed_cut.supervisions_feature_mask()
ones = np.index_exp[list(
chain(range(0, 50), range(60, 80), range(200, 250),
range(260, 280)))]
zeros = np.index_exp[list(
chain(range(50, 60), range(80, 200), range(250, 260),
range(280, 400)))]
assert (mask[ones] == 1).all()
assert (mask[zeros] == 0).all()
def make_cut(sampling_rate: int, num_samples: int) -> Cut:
with make_recording(sampling_rate, num_samples) as recording:
duration = num_samples / sampling_rate
yield Cut(id=f'cut-{sampling_rate}-{duration}',
start=0,
duration=duration,
channel=0,
recording=recording)
def random_cut_set(n_cuts=100) -> CutSet:
return CutSet.from_cuts(
Cut(id=uuid4(),
start=round(random.uniform(0, 5), ndigits=8),
duration=round(random.uniform(3, 10), ndigits=8),
channel=0,
recording=Recording(id=uuid4(),
sources=[],
sampling_rate=16000,
num_samples=1600000,
duration=100.0)) for _ in range(n_cuts))
def cut_with_supervision_start01(recording):
return Cut(id='cut',
start=0.1,
duration=0.4,
channel=0,
supervisions=[
SupervisionSegment(id='sup',
recording_id='rec',
start=0.1,
duration=0.3)
],
recording=recording)
def test_augmentation_chain_randomized(target_sampling_rate: int,
sp_factor: float, resample_first: bool,
cut_duration: Seconds):
recording = Recording.from_file(
'test/fixtures/libri/libri-1088-134315-0000.wav')
if resample_first:
recording_aug = recording.resample(target_sampling_rate).perturb_speed(
sp_factor)
else:
recording_aug = recording.perturb_speed(sp_factor).resample(
target_sampling_rate)
audio_aug = recording_aug.load_audio()
assert audio_aug.shape[1] == recording_aug.num_samples
cut_aug = Cut(id='dummy',
start=0.5125,
duration=cut_duration,
channel=0,
recording=recording_aug)
assert cut_aug.load_audio().shape[1] == cut_aug.num_samples
def with_cut(self,
sampling_rate: int,
num_samples: int,
features: bool = True) -> Cut:
duration = num_samples / sampling_rate
cut = Cut(id=str(uuid4()),
start=0,
duration=duration,
channel=0,
recording=self.with_recording(sampling_rate=sampling_rate,
num_samples=num_samples))
if features:
cut = self._with_features(cut)
return cut
def deserialize_item(data: dict) -> Any:
# Figures out what type of manifest is being decoded with some heuristics
# and returns a Lhotse manifest object rather than a raw dict.
from lhotse import Cut, Features, Recording, SupervisionSegment
from lhotse.cut import MixedCut
data = arr2list_recursive(data)
if 'sources' in data:
return Recording.from_dict(data)
if 'num_features' in data:
return Features.from_dict(data)
if 'type' not in data:
return SupervisionSegment.from_dict(data)
cut_type = data.pop('type')
if cut_type == 'Cut':
return Cut.from_dict(data)
if cut_type == 'MixedCut':
return MixedCut.from_dict(data)
raise ValueError(f"Unexpected cut type during deserialization: '{cut_type}'")
def cut_set():
cut = Cut(id='cut-1',
start=0.0,
duration=10.0,
channel=0,
features=Features(
type='fbank',
num_frames=100,
num_features=40,
frame_shift=0.01,
sampling_rate=16000,
start=0.0,
duration=10.0,
storage_type='lilcom',
storage_path='irrelevant',
storage_key='irrelevant',
),
recording=Recording(id='rec-1',
sampling_rate=16000,
num_samples=160000,
duration=10.0,
sources=[
AudioSource(type='file',
channels=[0],
source='irrelevant')
]),
supervisions=[
SupervisionSegment(id='sup-1',
recording_id='irrelevant',
start=0.5,
duration=6.0),
SupervisionSegment(id='sup-2',
recording_id='irrelevant',
start=7.0,
duration=2.0)
])
return CutSet.from_cuts([
cut,
fastcopy(cut, id='cut-nosup', supervisions=[]),
fastcopy(cut, id='cut-norec', recording=None),
fastcopy(cut, id='cut-nofeat', features=None),
cut.pad(duration=30.0, direction='left'),
cut.pad(duration=30.0, direction='right'),
cut.pad(duration=30.0, direction='both'),
cut.mix(cut, offset_other_by=5.0, snr=8)
])
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)
def cut(recording):
return Cut(id='cut', start=0, duration=1.0, channel=0, recording=recording)
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 random_cut_set(n_cuts=100) -> CutSet:
return CutSet.from_cuts(
Cut(id=uuid4(),
start=round(random.uniform(0, 5), ndigits=8),
duration=round(random.uniform(3, 10), ndigits=8),
channel=0) for _ in range(n_cuts))
