def test_augmentation_chain_randomized(
recording: Recording,
rir: Recording,
target_sampling_rate: int,
sp_factor: float,
vp_factor: float,
reverb: bool,
resample_first: bool,
cut_duration: Seconds,
):
if resample_first:
recording_aug = (recording.resample(target_sampling_rate).
perturb_speed(sp_factor).perturb_volume(vp_factor))
else:
recording_aug = (recording.perturb_speed(sp_factor).resample(
target_sampling_rate).perturb_volume(vp_factor))
if reverb:
recording_aug = recording_aug.reverb_rir(rir)
audio_aug = recording_aug.load_audio()
assert audio_aug.shape[1] == recording_aug.num_samples
cut_aug = MonoCut(
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 mixed_overlapping_cut_set():
"""
Input mixed cut::
|---------------mixedcut--------------------|
|--------rec1 0-30s--------|
|-------rec2 15-45s--------|
|---sup1--| |-----sup3-----|
|sup2|
"""
cut_set = CutSet.from_cuts([
MonoCut(
'cut1', start=0, duration=30, channel=0,
recording=Recording(
id='rec1', sources=[], sampling_rate=16000, num_samples=160000, duration=60.0
),
supervisions=[
SupervisionSegment('sup1', 'rec1', start=1.5, duration=10.5),
SupervisionSegment('sup2', 'rec1', start=10, duration=6),
]
).mix(
MonoCut(
'cut2', start=15, duration=30, channel=0,
recording=Recording(
id='rec2', sources=[], sampling_rate=16000, num_samples=160000, duration=60.0
),
supervisions=[
SupervisionSegment('sup3', 'rec2', start=8, duration=18),
]
),
offset_other_by=15.0
)
])
assert isinstance(cut_set[0], MixedCut)
return cut_set
def test_trim_to_supervisions_mixed_cuts():
cut_set = CutSet.from_cuts([
Cut('cut1',
start=0,
duration=30,
channel=0,
recording=Recording(id='rec1',
sources=[],
sampling_rate=16000,
num_samples=160000,
duration=10.0),
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,
recording=Recording(id='rec1',
sources=[],
sampling_rate=16000,
num_samples=160000,
duration=10.0),
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_resample_opus():
# Test that reading OPUS files after resampling
# does not raise an exception.
r = Recording.from_file("test/fixtures/mono_c0.opus")
r.load_audio()
r1 = r.resample(24000)
r1.load_audio()
def test_audio_caching_disabled_works():
lhotse.set_caching_enabled(False) # Disable caching.
np.random.seed(89) # Reproducibility.
# Prepare two different waveforms.
noise1 = np.random.rand(1, 32000).astype(np.float32)
noise2 = np.random.rand(1, 32000).astype(np.float32)
# Sanity check -- the noises are different
assert np.abs(noise1 - noise2).sum() != 0
# Save the first waveform in a file.
with NamedTemporaryFile(suffix=".wav") as f:
torchaudio.save(f.name, torch.from_numpy(noise1), sample_rate=16000)
recording = Recording.from_file(f.name)
# Read the audio -- should be equal to noise1.
audio = recording.load_audio()
np.testing.assert_almost_equal(audio, noise1)
# Save noise2 to the same location.
torchaudio.save(f.name, torch.from_numpy(noise2), sample_rate=16000)
# Read the audio -- should be equal to noise2,
# and the caching is ignored (doesn't happen).
audio = recording.load_audio()
np.testing.assert_almost_equal(audio, noise2)
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 _process_file(
file_path: Pathlike,
speaker_metadata: Dict[str, SpeakerMetadata],
) -> Tuple[Recording, SupervisionSegment]:
"""
Process a single wav file and return a Recording and a SupervisionSegment.
"""
speaker_id = file_path.parent.parent.stem
session_id = file_path.parent.stem
uttid = file_path.stem
recording_id = f"{speaker_id}-{session_id}-{uttid}"
recording = Recording.from_file(file_path, recording_id=recording_id)
supervision = SupervisionSegment(
id=recording_id,
recording_id=recording_id,
speaker=speaker_id,
gender=speaker_metadata[speaker_id].gender,
start=0.0,
duration=recording.duration,
custom={
"speaker_name": speaker_metadata[speaker_id].name,
"nationality": speaker_metadata[speaker_id].nationality,
"split": speaker_metadata[speaker_id].split,
},
)
return recording, supervision
def test_cut_load_custom_recording_pad_left():
sampling_rate = 16000
duration = 52.4
audio = np.random.randn(1, compute_num_samples(
duration, sampling_rate)).astype(np.float32)
audio /= np.abs(audio).max() # normalize to [-1, 1]
with NamedTemporaryFile(suffix=".wav") as f:
torchaudio.save(f.name, torch.from_numpy(audio), sampling_rate)
f.flush()
os.fsync(f)
recording = Recording.from_file(f.name)
# Note: MonoCut doesn't normally have an "alignment" attribute,
# and a "load_alignment()" method.
# We are dynamically extending it.
cut = MonoCut(
id="x",
start=0,
duration=duration,
channel=0,
recording=dummy_recording(0, duration=duration),
)
cut.my_favorite_song = recording
cut_pad = cut.pad(duration=60.0, direction="left")
restored_audio = cut_pad.load_my_favorite_song()
assert restored_audio.shape == (1, 960000) # 16000 * 60
np.testing.assert_almost_equal(0, restored_audio[:, :-audio.shape[1]])
np.testing.assert_almost_equal(audio, restored_audio[:,
-audio.shape[1]:])
def prepare_tedlium(
tedlium_root: Pathlike,
output_dir: Optional[Pathlike] = None
) -> Dict[str, Dict[str, Union[RecordingSet, SupervisionSet]]]:
"""
Prepare manifests for the TED-LIUM v3 corpus.
The manifests are created in a dict with three splits: train, dev and test.
Each split contains a RecordingSet and SupervisionSet in a dict under keys 'recordings' and 'supervisions'.
:param tedlium_root: Path to the unpacked TED-LIUM data.
:return: A dict with standard corpus splits containing the manifests.
"""
tedlium_root = Path(tedlium_root)
output_dir = Path(output_dir) if output_dir is not None else None
corpus = {}
for split in ("train", "dev", "test"):
root = tedlium_root / "legacy" / split
recordings = RecordingSet.from_recordings(
Recording.from_file(p) for p in (root / "sph").glob("*.sph"))
stms = list((root / "stm").glob("*.stm"))
assert len(stms) == len(recordings), (
f"Mismatch: found {len(recordings)} "
f"sphere files and {len(stms)} STM files. "
f"You might be missing some parts of TEDLIUM...")
segments = []
for p in stms:
with p.open() as f:
for idx, l in enumerate(f):
rec_id, _, _, start, end, _, *words = l.split()
start, end = float(start), float(end)
text = " ".join(words).replace("{NOISE}", "[NOISE]")
if text == "ignore_time_segment_in_scoring":
continue
segments.append(
SupervisionSegment(
id=f"{rec_id}-{idx}",
recording_id=rec_id,
start=start,
duration=round(end - start, ndigits=8),
channel=0,
text=text,
language="English",
speaker=rec_id,
))
supervisions = SupervisionSet.from_segments(segments)
corpus[split] = {
"recordings": recordings,
"supervisions": supervisions
}
validate_recordings_and_supervisions(**corpus[split])
if output_dir is not None:
recordings.to_file(output_dir /
f"tedlium_recordings_{split}.jsonl.gz")
supervisions.to_file(output_dir /
f"tedlium_supervisions_{split}.jsonl.gz")
return corpus
def prepare_vctk(
corpus_dir: Pathlike,
output_dir: Optional[Pathlike] = None,
) -> Dict[str, Union[RecordingSet, SupervisionSet]]:
"""
Prepares and returns the L2 Arctic manifests which consist of Recordings and Supervisions.
:param corpus_dir: Pathlike, the path of the data dir.
:param output_dir: Pathlike, the path where to write the manifests.
:return: a dict with keys "read" and "spontaneous".
Each hold another dict of {'recordings': ..., 'supervisions': ...}
"""
corpus_dir = Path(corpus_dir)
assert corpus_dir.is_dir(), f"No such directory: {corpus_dir}"
speaker_meta = _parse_speaker_description(corpus_dir)
recordings = RecordingSet.from_recordings(
Recording.from_file(wav)
for wav in (corpus_dir / "wav48").rglob("*.wav"))
supervisions = []
for path in (corpus_dir / "txt").rglob("*.txt"):
# One utterance (line) per file
text = path.read_text().strip()
speaker = path.name.split("_")[0] # p226_001.txt -> p226
seg_id = path.stem
meta = speaker_meta.get(speaker, defaultdict(lambda: None))
if meta is None:
logging.warning(f"Cannot find metadata for speaker {speaker}.")
supervisions.append(
SupervisionSegment(
id=seg_id,
recording_id=seg_id,
start=0,
duration=recordings[seg_id].duration,
text=text,
language="English",
speaker=speaker,
gender=meta["gender"],
custom={
"accent": meta["accent"],
"age": meta["age"],
"region": meta["region"],
},
))
supervisions = SupervisionSet.from_segments(supervisions)
# note(pzelasko): There were 172 recordings without supervisions when I ran it.
# I am just removing them.
recordings, supervisions = remove_missing_recordings_and_supervisions(
recordings, supervisions)
validate_recordings_and_supervisions(recordings, supervisions)
if output_dir is not None:
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
recordings.to_json(output_dir / "recordings.json")
supervisions.to_json(output_dir / "supervisions.json")
return {"recordings": recordings, "supervisions": supervisions}
def process_file(
file_path: Any,
transcript_dict : Dict
) -> Tuple[Recording, SupervisionSegment]:
"""
Process a single wav file and return a Recording and a SupervisionSegment.
"""
try:
speaker_id = Path(file_path).stem
uttid = speaker_id
recording_id = speaker_id
recording = Recording.from_file(file_path)
text = transcript_dict[uttid]
supervision = SupervisionSegment(
id=recording_id,
recording_id=recording_id,
speaker=speaker_id,
start=0.0,
duration=recording.duration,
channel=0,
language="Chinese",
text=text.strip(),
)
return recording, supervision
except Exception as e:
logging.info(f"process_file err: {file_path}")
return None, None
def prepare_single_partition(
raw_manifest_path: Path,
corpus_dir: Path,
speaker_id: str,
clean_or_other: str,
):
recordings = []
supervisions = []
for meta in load_jsonl(raw_manifest_path):
recording = Recording.from_file(corpus_dir / meta["audio_filepath"])
recordings.append(recording)
supervisions.append(
SupervisionSegment(
id=recording.id,
recording_id=recording.id,
start=0,
duration=recording.duration,
channel=0,
text=meta["text"],
speaker=ID2SPEAKER[speaker_id],
gender=ID2GENDER[speaker_id],
custom={
"text_punct": meta["text_normalized"],
"split": clean_or_other
},
))
recordings = RecordingSet.from_recordings(recordings)
supervisions = SupervisionSet.from_segments(supervisions)
validate_recordings_and_supervisions(recordings, supervisions)
return recordings, supervisions
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 prepare_tedlium(
tedlium_root: Pathlike,
output_dir: Optional[Pathlike] = None
) -> Dict[str, Dict[str, Union[RecordingSet, SupervisionSet]]]:
"""
Prepare manifests for the TED-LIUM v3 corpus.
The manifests are created in a dict with three splits: train, dev and test.
Each split contains a RecordingSet and SupervisionSet in a dict under keys 'recordings' and 'supervisions'.
:param tedlium_root: Path to the unpacked TED-LIUM data.
:return: A dict with standard corpus splits containing the manifests.
"""
tedlium_root = Path(tedlium_root)
output_dir = Path(output_dir) if output_dir is not None else None
corpus = {}
for split in ('train', 'dev', 'test'):
root = tedlium_root / 'legacy' / split
recordings = RecordingSet.from_recordings(
Recording.from_file(p) for p in (root / 'sph').glob('*.sph')
)
stms = list((root / 'stm').glob('*.stm'))
assert len(stms) == len(recordings), f'Mismatch: found {len(recordings)} ' \
f'sphere files and {len(stms)} STM files. ' \
f'You might be missing some parts of TEDLIUM...'
segments = []
for p in stms:
with p.open() as f:
for idx, l in enumerate(f):
rec_id, _, _, start, end, _, *words = l.split()
start, end = float(start), float(end)
text = ' '.join(words).replace('{NOISE}', '[NOISE]')
if text == 'ignore_time_segment_in_scoring':
continue
segments.append(
SupervisionSegment(
id=f'{rec_id}-{idx}',
recording_id=rec_id,
start=start,
duration=round(end - start, ndigits=8),
channel=0,
text=text,
language='English',
speaker=rec_id,
)
)
supervisions = SupervisionSet.from_segments(segments)
corpus[split] = {
'recordings': recordings,
'supervisions': supervisions
}
validate_recordings_and_supervisions(**corpus[split])
if output_dir is not None:
recordings.to_json(output_dir / f'{split}_recordings.json')
supervisions.to_json(output_dir / f'{split}_supervisions.json')
return corpus
def test_cut_with_audio_move_to_memory():
path = "test/fixtures/mono_c0.wav"
cut = dummy_cut(0, duration=0.5).drop_recording()
cut.recording = Recording.from_file(path)
memory_cut = cut.move_to_memory()
np.testing.assert_equal(memory_cut.load_audio(), cut.load_audio())
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 recording():
return Recording(
id='rec',
sources=[AudioSource(type='file', channels=[0, 1], source='test/fixtures/stereo.wav')],
sampling_rate=8000,
num_samples=8000,
duration=1.0
)
def recording(file_source):
return Recording(
id="rec",
sources=[file_source],
sampling_rate=8000,
num_samples=4000,
duration=0.5,
)
def prepare_single_babel_language(corpus_dir: Pathlike, output_dir: Optional[Pathlike] = None):
manifests = defaultdict(dict)
for split in ('dev', 'eval', 'training'):
audio_dir = corpus_dir / f'conversational/{split}/audio'
recordings = RecordingSet.from_recordings(Recording.from_sphere(p) for p in audio_dir.glob('*.sph'))
if len(recordings) == 0:
logging.warning(f"No SPHERE files found in {audio_dir}")
manifests[split]['recordings'] = recordings
supervisions = []
text_dir = corpus_dir / f'conversational/{split}/transcription'
for p in text_dir.glob('*'):
# p.stem -> BABEL_BP_101_10033_20111024_205740_inLine
# parts:
# 0 -> BABEL
# 1 -> BP
# 2 -> <language-code> (101)
# 3 -> <speaker-id> (10033)
# 4 -> <date> (20111024)
# 5 -> <hour> (205740)
# 6 -> channel (inLine) ; inLine <=> A ; outLine <=> B ; "scripted" <=> A
p0, p1, lang_code, speaker, date, hour, channel, *_ = p.stem.split('_')
channel = {'inLine': 'A', 'outLine': 'B'}.get(channel, 'A')
# Add a None at the end so that the last timestamp is only used as "next_timestamp"
# and ends the iretation (otherwise we'd lose the last segment).
lines = p.read_text().splitlines() + [None]
for (timestamp, text), (next_timestamp, _) in sliding_window(2, zip(lines[::2], lines[1::2])):
start = float(timestamp[1:-1])
end = float(next_timestamp[1:-1])
supervisions.append(
SupervisionSegment(
id=f'{lang_code}_{speaker}_{channel}_{date}_{hour}_{int(100 * start):06}',
recording_id=p.stem,
start=start,
duration=round(end - start, ndigits=8),
channel=0,
text=normalize_text(text),
language=BABELCODE2LANG[lang_code],
speaker=speaker,
)
)
if len(supervisions) == 0:
logging.warning(f"No supervisions found in {text_dir}")
manifests[split]['supervisions'] = SupervisionSet.from_segments(supervisions)
validate_recordings_and_supervisions(
manifests[split]['recordings'],
manifests[split]['superevisions']
)
if output_dir is not None:
language = BABELCODE2LANG[lang_code]
if split == 'training':
split = 'train'
manifests[split]['recordings'].to_json(f'recordings_{language}_{split}.json')
manifests[split]['supervisions'].to_json(f'supervisions_{language}_{split}.json')
return manifests
def prepare_cmu_arctic(
corpus_dir: Pathlike,
output_dir: Optional[Pathlike] = None,
) -> Dict[str, Union[RecordingSet, SupervisionSet]]:
"""
Prepares and returns the CMU Arctic manifests,
which consist of Recordings and Supervisions.
:param corpus_dir: Pathlike, the path of the data dir.
:param output_dir: Pathlike, the path where to write the manifests.
:return: a dict of {'recordings': ..., 'supervisions': ...}
"""
corpus_dir = Path(corpus_dir)
assert corpus_dir.is_dir(), f"No such directory: {corpus_dir}"
recordings = RecordingSet.from_recordings(
# Example ID: cmu_us_sup_arctic-arctic_a0001
Recording.from_file(
wav, recording_id=f"{_get_speaker(wav.parent.parent.name)}-{wav.stem}"
)
for wav in corpus_dir.rglob("*.wav")
)
supervisions = []
for path in corpus_dir.rglob("txt.done.data"):
lines = path.read_text().splitlines()
speaker = _get_speaker(path.parent.parent.name)
for l in lines:
l = l[2:-2] # get rid of parentheses and whitespaces on the edges
seg_id, text = l.split(maxsplit=1)
seg_id = f"{speaker}-{seg_id}"
supervisions.append(
SupervisionSegment(
id=seg_id,
recording_id=seg_id,
start=0,
duration=recordings[seg_id].duration,
text=text.replace('"', ""), # get rid of quotation marks,
language="English",
speaker=speaker,
gender=GENDER_MAP.get(speaker),
custom={"accent": ACCENT_MAP.get(speaker)},
)
)
supervisions = SupervisionSet.from_segments(supervisions)
# There seem to be 20 recordings missing; remove the before validation
recordings, supervisions = remove_missing_recordings_and_supervisions(
recordings, supervisions
)
validate_recordings_and_supervisions(recordings, supervisions)
if output_dir is not None:
output_dir = Path(output_dir)
recordings.to_json(output_dir / "cmu_arctic_recordings.json")
supervisions.to_json(output_dir / "cmu_arctic_supervisions.json")
return {"recordings": recordings, "supervisions": supervisions}
def __post_init__(self):
if isinstance(self.rir, dict):
from lhotse import Recording
# Pass a shallow copy of the RIR dict since `from_dict()` pops the `sources` key.
self.rir = Recording.from_dict(self.rir.copy())
assert all(
c < self.rir.num_channels for c in
self.rir_channels), "Invalid channel index in `rir_channels`"
def test_cut_with_audio_move_to_memory_large_offset():
path = "test/fixtures/mono_c0.wav"
cut = dummy_cut(0, duration=0.1).drop_recording()
cut.recording = Recording.from_file(path)
cut.start = 0.4
assert isclose(cut.end, 0.5)
memory_cut = cut.move_to_memory()
np.testing.assert_equal(memory_cut.load_audio(), cut.load_audio())
def random_cut_set(n_cuts=100) -> CutSet:
return CutSet.from_cuts(
MonoCut(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 test_cut_trim_to_supervisions_extend_handles_end_of_recording(mono_cut):
"""
Scenario::
|----------Recording---------|
|---Sup1----| |--Sup2--|
|------------Cut-------------|
Into::
|----------Recording---------|
|---Cut1----| |---Cut2---|
|---Sup1----| |--Sup2--|
"""
cut = MonoCut(
id="X",
start=0.0,
duration=10.0,
channel=0,
supervisions=[
SupervisionSegment(id="X",
recording_id="X",
start=0.0,
duration=4.0),
SupervisionSegment(id="X",
recording_id="X",
start=7.0,
duration=3.0),
],
recording=Recording(id="X",
sources=[],
sampling_rate=8000,
num_samples=80000,
duration=10.0),
)
cuts = cut.trim_to_supervisions(min_duration=4.0)
assert len(cuts) == 2
c1, c2 = cuts
assert c1.start == 0
assert c1.duration == 4.0
assert len(c1.supervisions) == 1
(c1_s1, ) = c1.supervisions
assert c1_s1.start == 0.0
assert c1_s1.duration == 4.0
assert c2.start == 6.5
assert c2.duration == 3.5
assert len(c2.supervisions) == 1
(c2_s1, ) = c2.supervisions
assert c2_s1.start == 0.5
assert c2_s1.duration == 3.0
def recording():
return Recording(
id="rec",
sources=[
AudioSource(type="file",
channels=[0, 1],
source="test/fixtures/stereo.wav")
],
sampling_rate=8000,
num_samples=8000,
duration=1.0,
)
def with_recording(self, sampling_rate: int,
num_samples: int) -> Recording:
f = NamedTemporaryFile('wb', suffix='.wav')
self.files.append(f)
duration = num_samples / sampling_rate
samples = np.random.rand(num_samples)
soundfile.write(f.name, samples, samplerate=sampling_rate)
return Recording(
id=str(uuid4()),
sources=[AudioSource(type='file', channels=[0], source=f.name)],
sampling_rate=sampling_rate,
num_samples=num_samples,
duration=duration)
def make_recording(sampling_rate: int, num_samples: int) -> Recording:
# The idea is that we're going to write to a temporary file with a sine wave recording
# of specified duration and sampling rate, and clean up only after the test is executed.
with NamedTemporaryFile('wb', suffix='.wav') as f:
duration = num_samples / sampling_rate
samples: np.ndarray = np.sin(2 * np.pi * np.arange(0, num_samples) /
sampling_rate)
soundfile.write(f, samples, samplerate=sampling_rate)
yield Recording(
id=f'recording-{sampling_rate}-{duration}',
sources=[AudioSource(type='file', channels=[0], source=f.name)],
sampling_rate=sampling_rate,
num_samples=num_samples,
duration=duration)
def test_cut_move_to_memory_audio_serialization():
path = "test/fixtures/mono_c0.wav"
cut = dummy_cut(0, duration=0.5).drop_recording()
cut.recording = Recording.from_file(path)
cut_with_audio = cut.move_to_memory()
assert cut.custom is None # original cut is unmodified
data = cut_with_audio.to_dict()
cut_deserialized = MonoCut.from_dict(data)
np.testing.assert_equal(cut_deserialized.load_audio(),
cut_with_audio.load_audio())
def recording_set():
return RecordingSet.from_recordings([
Recording(id='x',
sources=[
AudioSource(type='file',
channels=[0],
source='text/fixtures/mono_c0.wav'),
AudioSource(type='command',
channels=[1],
source='cat text/fixtures/mono_c1.wav')
],
sampling_rate=8000,
num_samples=4000,
duration=0.5)
])
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),
])
