def parse_utterance(
audio: Any, root_path: Path, subsets: Sequence
) -> Tuple[Recording, Dict[str, List[SupervisionSegment]]]:
sampling_rate = 16000
recording = Recording(
id=audio["aid"],
sources=[
AudioSource(
type="file",
channels=[0],
source=str(root_path / audio["path"]),
)
],
num_samples=compute_num_samples(duration=audio["duration"],
sampling_rate=sampling_rate),
sampling_rate=sampling_rate,
duration=audio["duration"],
)
segments = defaultdict(dict)
for sub in subsets:
segments[sub] = []
for seg in audio["segments"]:
segment = SupervisionSegment(
id=seg["sid"],
recording_id=audio["aid"],
start=seg["begin_time"],
duration=add_durations(seg["end_time"], -seg["begin_time"],
sampling_rate),
language="Chinese",
text=seg["text"].strip(),
)
for sub in seg["subsets"]:
if sub in subsets:
segments[sub].append(segment)
return recording, segments
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 parse_utterance(
audio: Any, root_path: Path
) -> Optional[Tuple[Recording, List[SupervisionSegment]]]:
sampling_rate = int(audio["sample_rate"])
recording = Recording(
id=audio["aid"],
sources=[
AudioSource(
type="file",
channels=list(range(int(audio["channels"]))),
source=str(root_path / audio["path"]),
)
],
num_samples=compute_num_samples(
duration=Seconds(audio["duration"]), sampling_rate=sampling_rate
),
sampling_rate=sampling_rate,
duration=Seconds(audio["duration"]),
)
segments = []
for seg in audio["segments"]:
segments.append(
SupervisionSegment(
id=seg["sid"],
recording_id=audio["aid"],
start=Seconds(seg["begin_time"]),
duration=round(Seconds(seg["end_time"] - seg["begin_time"]), ndigits=8),
channel=0,
language="English",
speaker=seg["speaker"],
text=seg["text_tn"],
)
)
return recording, segments
def test_cut_load_custom_recording_truncate():
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 = dummy_cut(0, duration=duration)
cut.my_favorite_song = recording
cut_trunc = cut.truncate(duration=5.0)
restored_audio = cut_trunc.load_my_favorite_song()
assert restored_audio.shape == (1, 80000)
np.testing.assert_almost_equal(audio[:, :80000], restored_audio)
