def test_trim_to_supervisions_mixed_cuts():
cut_set = CutSet.from_cuts([
Cut('cut1',
start=0,
duration=30,
channel=0,
supervisions=[
SupervisionSegment('sup1', 'rec1', start=1.5, duration=8.5),
SupervisionSegment('sup2', 'rec1', start=10, duration=5),
SupervisionSegment('sup3', 'rec1', start=20, duration=8),
]).append(
Cut('cut2',
start=0,
duration=30,
channel=0,
supervisions=[
SupervisionSegment('sup4',
'rec1',
start=0,
duration=30),
]))
])
assert isinstance(cut_set[0], MixedCut)
cuts = cut_set.trim_to_supervisions()
assert len(cuts) == 4
# After "trimming", the MixedCut "decayed" into simple, unmixed cuts, as they did not overlap
assert all(isinstance(cut, Cut) for cut in cuts)
assert all(len(cut.supervisions) == 1 for cut in cuts)
assert all(cut.supervisions[0].start == 0 for cut in cuts)
cut = cuts[0]
# Check that the cuts preserved their start/duration/supervisions after trimming
assert cut.start == 1.5
assert cut.duration == 8.5
assert cut.supervisions[0].id == 'sup1'
cut = cuts[1]
assert cut.start == 10
assert cut.duration == 5
assert cut.supervisions[0].id == 'sup2'
cut = cuts[2]
assert cut.start == 20
assert cut.duration == 8
assert cut.supervisions[0].id == 'sup3'
cut = cuts[3]
assert cut.start == 0
assert cut.duration == 30
assert cut.supervisions[0].id == 'sup4'
def test_mix_same_recording_channels():
recording = Recording('rec', sampling_rate=8000, num_samples=30 * 8000, duration=30, sources=[
AudioSource('file', channels=[0], source='irrelevant1.wav'),
AudioSource('file', channels=[1], source='irrelevant2.wav')
])
cut_set = CutSet.from_cuts([
Cut('cut1', start=0, duration=30, channel=0, recording=recording),
Cut('cut2', start=0, duration=30, channel=1, recording=recording)
])
mixed = cut_set.mix_same_recording_channels()
assert len(mixed) == 1
cut = mixed[0]
assert isinstance(cut, MixedCut)
assert len(cut.tracks) == 2
assert cut.tracks[0].cut == cut_set[0]
assert cut.tracks[1].cut == cut_set[1]
def _read_audio(cut: Cut,
suppress_errors: bool = False) -> Optional[torch.Tensor]:
"""
Loads audio data from cut, or returns None if there was an error
and ``suppress_errors`` was set to ``True``.
"""
with suppress_and_warn(AudioLoadingError,
DurationMismatchError,
enabled=suppress_errors):
return torch.from_numpy(cut.load_audio()[0])
def cut2(dummy_features):
return Cut(id='cut-2',
start=180.0,
duration=10.0,
features=dummy_features,
supervisions=[
SupervisionSegment(id='sup-3',
recording_id='irrelevant',
start=3.0,
duration=2.5)
])
def _read_audio(
cut: Cut,
suppress_errors: bool = False,
recording_field: Optional[str] = None) -> Optional[torch.Tensor]:
"""
Loads audio data from cut, or returns None if there was an error
and ``suppress_errors`` was set to ``True``.
"""
with suppress_audio_loading_errors(enabled=suppress_errors):
if recording_field is None:
audio = cut.load_audio()
else:
attr = getattr(cut, recording_field)
assert isinstance(
attr, Recording
), f"Expected 'getattr(cut, {recording_field})' to yield Recording, got {type(attr)}"
audio = cut.load_custom(recording_field)
assert audio.shape[
0] == 1, f"Expected single-channel audio in cut:\n{cut}"
return torch.from_numpy(audio[0])
def dummy_cut(id: str = 'irrelevant',
start: float = 0.0,
duration: float = 1.0,
supervisions=None):
return Cut(
id=id,
start=start,
duration=duration,
features=dummy_features(0),
supervisions=supervisions if supervisions is not None else [],
)
def dummy_cut(unique_id: int,
start: float = 0.0,
duration: float = 1.0,
supervisions=None):
return Cut(
id=f'dummy-cut-{unique_id:04d}',
start=start,
duration=duration,
channel=0,
features=dummy_features(unique_id),
supervisions=supervisions if supervisions is not None else [],
)
def test_trim_to_supervisions_mixed_cuts():
cut_set = CutSet.from_cuts([
Cut('cut1',
start=0,
duration=30,
channel=0,
supervisions=[
SupervisionSegment('sup1', 'rec1', start=1.5, duration=8.5),
SupervisionSegment('sup2', 'rec1', start=10, duration=5),
SupervisionSegment('sup3', 'rec1', start=20, duration=8),
]).append(
Cut('cut2',
start=0,
duration=30,
channel=0,
supervisions=[
SupervisionSegment('sup4',
'rec1',
start=0,
duration=30),
]))
])
cuts = cut_set.trim_to_supervisions()
assert len(cuts) == 4
assert all(isinstance(cut, MixedCut) for cut in cuts)
assert all(cut.start == 0 for cut in cuts)
assert all(len(cut.supervisions) == 1 for cut in cuts)
assert all(cut.supervisions[0].start == 0 for cut in cuts)
cut = cuts[0]
assert cut.duration == 8.5
assert cut.supervisions[0].id == 'sup1'
cut = cuts[1]
assert cut.duration == 5
assert cut.supervisions[0].id == 'sup2'
cut = cuts[2]
assert cut.duration == 8
assert cut.supervisions[0].id == 'sup3'
cut = cuts[3]
assert cut.duration == 30
assert cut.supervisions[0].id == 'sup4'
def test_trim_to_unsupervised_segments():
cut_set = CutSet.from_cuts([
# Yields 3 unsupervised cuts - before first supervision,
# between sup2 and sup3, and after sup3.
Cut('cut1',
start=0,
duration=30,
channel=0,
supervisions=[
SupervisionSegment('sup1', 'rec1', start=1.5, duration=8.5),
SupervisionSegment('sup2', 'rec1', start=10, duration=5),
SupervisionSegment('sup3', 'rec1', start=20, duration=8),
]),
# Does not yield any "unsupervised" cut.
Cut('cut2',
start=0,
duration=30,
channel=0,
supervisions=[
SupervisionSegment('sup4', 'rec1', start=0, duration=30),
]),
])
unsupervised_cuts = cut_set.trim_to_unsupervised_segments()
assert len(unsupervised_cuts) == 3
assert unsupervised_cuts[0].start == 0
assert unsupervised_cuts[0].duration == 1.5
assert unsupervised_cuts[0].supervisions == []
assert unsupervised_cuts[1].start == 15
assert unsupervised_cuts[1].duration == 5
assert unsupervised_cuts[1].supervisions == []
assert unsupervised_cuts[2].start == 28
assert unsupervised_cuts[2].duration == 2
assert unsupervised_cuts[2].supervisions == []
def cut1(dummy_features):
return Cut(id='cut-1',
start=0.0,
duration=10.0,
features=dummy_features,
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)
])
def cut_with_relative_paths():
return Cut('cut',
0,
10,
0,
features=Features(type='fbank',
num_frames=1000,
num_features=40,
sampling_rate=8000,
storage_type='lilcom_files',
storage_path='storage_dir',
storage_key='feats.llc',
start=0,
duration=10),
recording=Recording('rec',
[AudioSource('file', [0], 'audio.wav')],
8000, 80000, 10.0))
def overlapping_supervisions_cut():
return Cut(
id='cut-1',
start=0.0,
duration=0.5,
channel=0,
features=Features(
recording_id='recording-1',
channels=0,
start=0,
duration=0.5,
type='fbank',
num_frames=50,
num_features=80,
sampling_rate=16000,
storage_type='lilcom',
storage_path=
'test/fixtures/dummy_feats/storage/e66b6386-aee5-4a5a-8369-fdde1d2b97c7.llc'
),
supervisions=[
SupervisionSegment(id='s1',
recording_id='recording-1',
start=0.0,
duration=0.2),
SupervisionSegment(id='s2',
recording_id='recording-1',
start=0.1,
duration=0.2),
SupervisionSegment(id='s3',
recording_id='recording-1',
start=0.2,
duration=0.2),
SupervisionSegment(id='s4',
recording_id='recording-1',
start=0.3,
duration=0.2)
])
def _read_features(cut: Cut) -> torch.Tensor:
return torch.from_numpy(cut.load_features())
def _read_audio(cut: Cut) -> torch.Tensor:
return torch.from_numpy(cut.load_audio()[0])
def validate_cut(c: Cut, read_data: bool = False) -> None:
# Validate MixedCut
if isinstance(c, MixedCut):
assert (len(c.tracks) >
0), f"MonoCut {c.id}: Mixed cut must have at least one track."
for idx, track in enumerate(c.tracks):
validate_cut(track.cut, read_data=read_data)
assert (track.offset >=
0), f"MonoCut: {c.id}: track {idx} has a negative offset."
return
# Validate MonoCut and PaddingCut
assert c.start >= 0, f"MonoCut {c.id}: start must be 0 or greater (got {c.start})"
assert (
c.duration > 0
), f"MonoCut {c.id}: duration must be greater than 0 (got {c.duration})"
assert (
c.sampling_rate > 0
), f"MonoCut {c.id}: sampling_rate must be greater than 0 (got {c.sampling_rate})"
assert (
c.has_features or c.has_recording
), f"MonoCut {c.id}: must have either Features or Recording attached."
# The rest pertains only to regular Cuts
if isinstance(c, PaddingCut):
return
# Conditions related to features
if c.has_features:
validate_features(c.features)
assert c.channel == c.features.channels
if read_data:
# We are not passing "read_data" to "validate_features" to avoid loading feats twice;
# we'll just validate the subset of the features relevant for the cut.
feats = c.load_features()
n_fr, n_ft = feats.shape
assert (
c.num_frames == n_fr
), f"MonoCut {c.id}: expected num_frames: {c.num_frames}, actual: {n_fr}"
assert (
c.num_features == n_ft
), f"MonoCut {c.id}: expected num_features: {c.num_features}, actual: {n_ft}"
# Conditions related to recording
if c.has_recording:
validate_recording(c.recording)
assert c.channel in c.recording.channel_ids
if read_data:
# We are not passing "read_data" to "validate_recording" to avoid loading audio twice;
# we'll just validate the subset of the recording relevant for the cut.
samples = c.load_audio()
assert (
c.num_samples == samples.shape[1]
), f"MonoCut {c.id}: expected {c.num_samples} samples, got {samples.shape[1]}"
# Conditions related to supervisions
for s in c.supervisions:
validate_supervision(s)
assert s.recording_id == c.recording_id, (
f"MonoCut {c.id}: supervision {s.id} has a mismatched recording_id "
f"(expected {c.recording_id}, supervision has {s.recording_id})")
assert s.channel == c.channel, (
f"MonoCut {c.id}: supervision {s.id} has a mismatched channel "
f"(expected {c.channel}, supervision has {s.channel})")
# Conditions related to custom fields
if c.custom is not None:
assert isinstance(
c.custom, dict
), f"MonoCut {c.id}: custom field has to be set to a dict or None."
for key, value in c.custom.items():
if isinstance(value, Array):
validate_array(value, read_data=read_data)
elif isinstance(value, TemporalArray):
validate_temporal_array(value, read_data=read_data)
if not isclose(c.duration, value.duration):
logging.warning(
f"MonoCut {c.id}: possibly mismatched "
f"duration between cut ({c.duration}s) and temporal array "
f"in custom field '{key}' (num_frames={value.num_frames} * "
f"frame_shift={value.frame_shift} == duration={value.duration})."
)
