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 Features, MonoCut, Recording, SupervisionSegment
from lhotse.array import deserialize_array
from lhotse.cut import MixedCut
if "shape" in data or "array" in data:
return deserialize_array(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 == "MonoCut":
return MonoCut.from_dict(data)
if cut_type == "Cut":
warnings.warn(
"Your manifest was created with Lhotse version earlier than v0.8, when MonoCut was called Cut. "
"Please re-generate it with Lhotse v0.8 as it might stop working in a future version "
"(using manifest.from_file() and then manifest.to_file() should be sufficient)."
)
return MonoCut.from_dict(data)
if cut_type == "MixedCut":
return MixedCut.from_dict(data)
raise ValueError(
f"Unexpected cut type during deserialization: '{cut_type}'")
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 _upload_one(item: Features, url: str) -> Features:
feats_mtx = item.load()
feats_writer = LilcomURLWriter(url)
new_key = feats_writer.write(key=item.storage_key, value=feats_mtx)
return fastcopy(item,
storage_path=url,
storage_key=new_key,
storage_type=feats_writer.name)
def test_validate_features_consistent_num_frames_does_not_raise():
manifest = Features(
type='irrelevant',
num_frames=100,
num_features=40,
frame_shift=0.01,
sampling_rate=16000,
start=0.0,
duration=1.0,
storage_type='irrelevant',
storage_path='irrelevant',
storage_key='irrelevant',
)
validate_features(manifest)
def test_validate_features_inconsistent_num_frames_raises():
manifest = Features(
type='irrelevant',
num_frames=101,
num_features=40,
frame_shift=0.01,
sampling_rate=16000,
start=0.0,
duration=1.0,
storage_type='irrelevant',
storage_path='irrelevant',
storage_key='irrelevant',
)
with pytest.raises(AssertionError):
validate_features(manifest)
def feature_set():
return FeatureSet(features=[
Features(recording_id='irrelevant',
channels=0,
start=0.0,
duration=20.0,
type='fbank',
num_frames=2000,
num_features=20,
frame_shift=0.01,
sampling_rate=16000,
storage_type='lilcom',
storage_path='/irrelevant/',
storage_key='path.llc')
])
def cut_set():
cut = MonoCut(
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 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 feature_set():
return FeatureSet(features=[
Features(
recording_id="irrelevant",
channels=0,
start=0.0,
duration=20.0,
type="fbank",
num_frames=2000,
num_features=20,
frame_shift=0.01,
sampling_rate=16000,
storage_type="lilcom",
storage_path="/irrelevant/",
storage_key="path.llc",
)
])
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_with_relative_paths():
return MonoCut(
"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,
frame_shift=0.01,
),
recording=Recording("rec", [AudioSource("file", [0], "audio.wav")],
8000, 80000, 10.0),
)
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 MonoCut, 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 == 'MonoCut':
return MonoCut.from_dict(data)
if cut_type == 'Cut':
warnings.warn('Your manifest was created with Lhotse version earlier than v0.8, when MonoCut was called Cut. '
'Please re-generate it with Lhotse v0.8 as it might stop working in a future version '
'(using manifest.from_file() and then manifest.to_file() should be sufficient).')
return MonoCut.from_dict(data)
if cut_type == 'MixedCut':
return MixedCut.from_dict(data)
raise ValueError(f"Unexpected cut type during deserialization: '{cut_type}'")
def load_kaldi_data_dir(
path: Pathlike,
sampling_rate: int,
frame_shift: Optional[Seconds] = None,
) -> Tuple[RecordingSet, Optional[SupervisionSet], Optional[FeatureSet]]:
"""
Load a Kaldi data directory and convert it to a Lhotse RecordingSet and SupervisionSet manifests.
For this to work, at least the wav.scp file must exist.
SupervisionSet is created only when a segments file exists.
All the other files (text, utt2spk, etc.) are optional, and some of them might not be handled yet.
In particular, feats.scp files are ignored.
"""
path = Path(path)
assert path.is_dir()
# must exist for RecordingSet
recordings = load_kaldi_text_mapping(path / 'wav.scp', must_exist=True)
durations = defaultdict(float)
reco2dur = path / 'reco2dur'
if not reco2dur.is_file():
raise ValueError(
f"No such file: '{reco2dur}' -- fix it by running: utils/data/get_reco2dur.sh <data-dir>"
)
with reco2dur.open() as f:
for line in f:
recording_id, dur = line.strip().split()
durations[recording_id] = float(dur)
recording_set = RecordingSet.from_recordings(
Recording(id=recording_id,
sources=[
AudioSource(type='command' if path_or_cmd.
endswith('|') else 'file',
channels=[0],
source=path_or_cmd[:-1] if path_or_cmd.
endswith('|') else path_or_cmd)
],
sampling_rate=sampling_rate,
num_samples=int(durations[recording_id] * sampling_rate),
duration=durations[recording_id])
for recording_id, path_or_cmd in recordings.items())
supervision_set = None
segments = path / 'segments'
if segments.is_file():
with segments.open() as f:
supervision_segments = [l.strip().split() for l in f]
texts = load_kaldi_text_mapping(path / 'text')
speakers = load_kaldi_text_mapping(path / 'utt2spk')
genders = load_kaldi_text_mapping(path / 'spk2gender')
languages = load_kaldi_text_mapping(path / 'utt2lang')
supervision_set = SupervisionSet.from_segments(
SupervisionSegment(id=segment_id,
recording_id=recording_id,
start=float(start),
duration=float(end) - float(start),
channel=0,
text=texts[segment_id],
language=languages[segment_id],
speaker=speakers[segment_id],
gender=genders[speakers[segment_id]])
for segment_id, recording_id, start, end in supervision_segments)
feature_set = None
feats_scp = path / 'feats.scp'
if feats_scp.exists() and is_module_available('kaldiio'):
if frame_shift is not None:
import kaldiio
from lhotse.features.io import KaldiReader
feature_set = FeatureSet.from_features(
Features(type='kaldiio',
num_frames=mat.shape[0],
num_features=mat.shape[1],
frame_shift=frame_shift,
sampling_rate=sampling_rate,
start=0,
duration=mat.shape[0] * frame_shift,
storage_type=KaldiReader.name,
storage_path=str(feats_scp),
storage_key=utt_id,
recording_id=supervision_set[utt_id].recording_id
if supervision_set is not None else utt_id,
channels=0)
for utt_id, mat in kaldiio.load_scp_sequential(str(feats_scp)))
else:
warnings.warn(f"Failed to import Kaldi 'feats.scp' to Lhotse: "
f"frame_shift must be not None. "
f"Feature import omitted.")
return recording_set, supervision_set, feature_set
def load_kaldi_data_dir(
path: Pathlike,
sampling_rate: int,
frame_shift: Optional[Seconds] = None,
map_string_to_underscores: Optional[str] = None,
num_jobs: int = 1,
) -> Tuple[RecordingSet, Optional[SupervisionSet], Optional[FeatureSet]]:
"""
Load a Kaldi data directory and convert it to a Lhotse RecordingSet and
SupervisionSet manifests. For this to work, at least the wav.scp file must exist.
SupervisionSet is created only when a segments file exists.
All the other files (text, utt2spk, etc.) are optional, and some of them might
not be handled yet. In particular, feats.scp files are ignored.
:param map_string_to_underscores: optional string, when specified, we will replace
all instances of this string in SupervisonSegment IDs to underscores.
This is to help with handling underscores in Kaldi
(see :func:`.export_to_kaldi`). This is also done for speaker IDs.
"""
path = Path(path)
assert path.is_dir()
def fix_id(t: str) -> str:
if map_string_to_underscores is None:
return t
return t.replace(map_string_to_underscores, "_")
# must exist for RecordingSet
recordings = load_kaldi_text_mapping(path / "wav.scp", must_exist=True)
with ProcessPoolExecutor(num_jobs) as ex:
dur_vals = ex.map(get_duration, recordings.values())
durations = dict(zip(recordings.keys(), dur_vals))
recording_set = RecordingSet.from_recordings(
Recording(
id=recording_id,
sources=[
AudioSource(
type="command" if path_or_cmd.endswith("|") else "file",
channels=[0],
source=path_or_cmd[:-1]
if path_or_cmd.endswith("|")
else path_or_cmd,
)
],
sampling_rate=sampling_rate,
num_samples=compute_num_samples(durations[recording_id], sampling_rate),
duration=durations[recording_id],
)
for recording_id, path_or_cmd in recordings.items()
)
supervision_set = None
segments = path / "segments"
if segments.is_file():
with segments.open() as f:
supervision_segments = [sup_string.strip().split() for sup_string in f]
texts = load_kaldi_text_mapping(path / "text")
speakers = load_kaldi_text_mapping(path / "utt2spk")
genders = load_kaldi_text_mapping(path / "spk2gender")
languages = load_kaldi_text_mapping(path / "utt2lang")
supervision_set = SupervisionSet.from_segments(
SupervisionSegment(
id=fix_id(segment_id),
recording_id=recording_id,
start=float(start),
duration=add_durations(
float(end), -float(start), sampling_rate=sampling_rate
),
channel=0,
text=texts[segment_id],
language=languages[segment_id],
speaker=fix_id(speakers[segment_id]),
gender=genders[speakers[segment_id]],
)
for segment_id, recording_id, start, end in supervision_segments
)
feature_set = None
feats_scp = path / "feats.scp"
if feats_scp.exists() and is_module_available("kaldi_native_io"):
if frame_shift is not None:
import kaldi_native_io
from lhotse.features.io import KaldiReader
feature_set = FeatureSet.from_features(
Features(
type="kaldi_native_io",
num_frames=mat_shape.num_rows,
num_features=mat_shape.num_cols,
frame_shift=frame_shift,
sampling_rate=sampling_rate,
start=0,
duration=mat_shape.num_rows * frame_shift,
storage_type=KaldiReader.name,
storage_path=str(feats_scp),
storage_key=utt_id,
recording_id=supervision_set[utt_id].recording_id
if supervision_set is not None
else utt_id,
channels=0,
)
for utt_id, mat_shape in kaldi_native_io.SequentialMatrixShapeReader(
f"scp:{feats_scp}"
)
)
else:
warnings.warn(
"Failed to import Kaldi 'feats.scp' to Lhotse: "
"frame_shift must be not None. "
"Feature import omitted."
)
return recording_set, supervision_set, feature_set