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,
NonPositiveEnergyError,
enabled=suppress_errors,
):
return torch.from_numpy(cut.load_audio()[0])
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 _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})."
)