def simple_mixed_cut():
return MixedCut(id='simple-mixed-cut',
tracks=[
MixTrack(cut=dummy_cut('cut0', duration=10.0)),
MixTrack(cut=dummy_cut('cut1', duration=10.0),
offset=5.0),
])
def simple_mixed_cut():
return MixedCut(
id="simple-mixed-cut",
tracks=[
MixTrack(cut=dummy_cut(0, duration=10.0)),
MixTrack(cut=dummy_cut(1, duration=10.0), offset=5.0),
],
)
def test_cut_set_extend_by():
cut1 = dummy_cut(int(uuid4()), start=0.0, duration=0.5)
cut2 = dummy_cut(int(uuid4()), start=0.2, duration=0.4)
cut_set = CutSet.from_cuts([cut1, cut2])
extended_cut_set = cut_set.extend_by(duration=0.3,
direction="both",
preserve_id=True)
assert isclose(extended_cut_set[cut1.id].start, 0.0)
assert isclose(extended_cut_set[cut1.id].end, 0.8)
assert isclose(extended_cut_set[cut2.id].start, 0.0)
assert isclose(extended_cut_set[cut2.id].end, 0.9)
def test_mixed_cut_fill_supervision_shrink():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut = cut.mix(dummy_cut(1)).truncate(
duration=0.5) # cuts are 100% overlapping
fcut = cut.fill_supervision(shrink_ok=True)
# Original is not modified
assert cut.supervisions[0].start == 0
assert cut.supervisions[0].duration == 1
# Result is modified
assert fcut.supervisions[0].start == 0
assert fcut.supervisions[0].duration == 0.5
def test_cut_pairs_sampler_2():
cut_set = CutSet.from_cuts([
dummy_cut(0, duration=10),
dummy_cut(1, duration=20),
])
sampler = CutPairsSampler(
source_cuts=cut_set,
target_cuts=cut_set,
max_source_duration=50,
max_target_duration=50,
)
batch = next(iter(sampler))
assert len(batch) == 2
def cutset():
return CutSet.from_cuts([
# MonoCut
dummy_cut(0, supervisions=[dummy_supervision(0)]),
# PaddingCut
PaddingCut('pad', duration=1.0, sampling_rate=16000, feat_value=-100,
num_frames=100, frame_shift=0.01, num_features=80, num_samples=16000),
# MixedCut
dummy_cut(0, supervisions=[dummy_supervision(0)]).mix(
dummy_cut(1, supervisions=[dummy_supervision(1)]),
offset_other_by=0.5,
snr=10
)
])
def test_token_collater(add_bos, add_eos):
test_sentences = [
"Testing the first sentence.",
"Let's add some more punctuation, shall we?",
"How about number 42!",
]
cuts = CutSet.from_cuts(
dummy_cut(idx,
idx,
supervisions=[dummy_supervision(idx, idx, text=sentence)])
for idx, sentence in enumerate(test_sentences))
token_collater = TokenCollater(cuts, add_bos=add_bos, add_eos=add_eos)
tokens_batch, tokens_lens = token_collater(cuts)
assert isinstance(tokens_batch, torch.LongTensor)
assert isinstance(tokens_lens, torch.IntTensor)
extend = int(add_bos) + int(add_eos)
expected_len = len(max(test_sentences, key=len)) + extend
assert tokens_batch.shape == (len(test_sentences), expected_len)
assert torch.all(tokens_lens == torch.IntTensor(
[len(sentence) + extend for sentence in test_sentences]))
reconstructed = token_collater.inverse(tokens_batch, tokens_lens)
assert reconstructed == test_sentences
def test_cut_set_decompose():
c = dummy_cut(
0,
start=5.0,
duration=10.0,
supervisions=[
dummy_supervision(0, start=0.0),
dummy_supervision(1, start=6.5)
],
)
assert c.start == 5.0
assert c.end == 15.0
cuts = CutSet.from_cuts([c])
recs, sups, feats = cuts.decompose()
assert isinstance(recs, RecordingSet)
assert len(recs) == 1
assert recs[0].id == "dummy-recording-0000"
assert isinstance(sups, SupervisionSet)
assert len(sups) == 2
assert sups[0].id == "dummy-segment-0000"
assert sups[0].start == 5.0
assert sups[0].end == 6.0
assert sups[1].id == "dummy-segment-0001"
assert sups[1].start == 11.5
assert sups[1].end == 12.5
assert isinstance(feats, FeatureSet)
assert len(feats) == 1
def dummy_cut_with_supervisions():
return dummy_cut(
unique_id=0,
supervisions=[
dummy_supervision(unique_id=i, duration=i) for i in range(1, 7)
],
)
def test_cut_split():
cut = dummy_cut(0, start=1.0, duration=4.0)
left, right = cut.split(2.0)
assert left.start == pytest.approx(1.0)
assert left.end == pytest.approx(3.0)
assert right.start == pytest.approx(3.0)
assert right.end == pytest.approx(5.0)
def test_extend_by_cut_with_supervision(
cut_start,
cut_duration,
extend_duration,
extend_direction,
supervision_start,
supervision_duration,
expected_start,
expected_end,
):
recording = dummy_recording(int(uuid4()), duration=1.0)
supervisions = SupervisionSet.from_segments([
SupervisionSegment(
id=int(uuid4()),
recording_id=recording.id,
start=supervision_start,
duration=supervision_duration,
)
])
cut = dummy_cut(int(uuid4()),
start=cut_start,
duration=cut_duration,
supervisions=supervisions)
extended_cut = cut.extend_by(duration=extend_duration,
direction=extend_direction)
assert isclose(extended_cut.supervisions[0].start, expected_start)
assert isclose(extended_cut.supervisions[0].end, expected_end)
def test_time_constraint_strictness():
normal = TimeConstraint(max_duration=100, strict=False)
strict = TimeConstraint(max_duration=100, strict=True)
# create cuts with large variance of durations
cut_durs = [30.0, 30.0, 10.0, 10.0, 20.0]
assert sum(cut_durs) == pytest.approx(100.0)
cuts = [dummy_cut(idx, duration=cd) for idx, cd in enumerate(cut_durs)]
# accumulate 80s of duration
for cut in cuts[:-1]:
normal.add(cut)
strict.add(cut)
assert normal.current == pytest.approx(80)
assert strict.current == pytest.approx(80)
# non-strict constraint is not close to exceeding (will accept next cut in a batch)
# strict constraint is close to exceeding (will not accept next cut in a batch)
assert not normal.close_to_exceeding()
assert strict.close_to_exceeding()
normal.add(cuts[-1])
strict.add(cuts[-1])
assert not normal.exceeded()
assert strict.exceeded()
def test_mixed_cut_fill_supervision_raises_on_two_supervisions():
cut = dummy_cut(0,
supervisions=[dummy_supervision(0),
dummy_supervision(1)])
cut = cut.pad(duration=10)
with pytest.raises(AssertionError):
fcut = cut.fill_supervision()
def test_mono_cut_fill_supervision_add_empty_true():
cut = dummy_cut(0)
fcut = cut.fill_supervision()
# Original is not modified
assert len(cut.supervisions) == 0
# Result is modified
assert fcut.supervisions[0].start == 0
assert fcut.supervisions[0].duration == 1
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 test_cut_set_decompose_output_dir_doesnt_duplicate_recording():
c = dummy_cut(0)
c2 = dummy_cut(0)
c2.id = "dummy-cut-0001" # override cut ID, retain identical recording ID as `c`
cuts = CutSet.from_cuts([c, c2])
with TemporaryDirectory() as td:
td = Path(td)
cuts.decompose(output_dir=td)
text = load_jsonl(td / "recordings.jsonl.gz")
print(list(text))
recs = load_manifest(td / "recordings.jsonl.gz")
assert isinstance(recs, RecordingSet)
# deduplicated recording
assert len(recs) == 1
assert recs[0].id == "dummy-recording-0000"
def test_extend_by_cut_preserve_id(preserve_id):
cut = dummy_cut(int(uuid4()), start=0.0, duration=0.5)
extended_cut = cut.extend_by(duration=0.3,
direction="right",
preserve_id=preserve_id)
if preserve_id:
assert extended_cut.id == cut.id
else:
assert extended_cut.id != cut.id
def test_cut_truncate_offset_with_nonzero_start():
cut = dummy_cut(0, start=1.0, duration=4.0)
left = cut.truncate(duration=2.0)
assert left.start == pytest.approx(1.0)
assert left.end == pytest.approx(3.0)
right = cut.truncate(offset=2.0)
assert right.start == pytest.approx(3.0)
assert right.end == pytest.approx(5.0)
def test_cut_set_decompose_doesnt_duplicate_recording():
c = dummy_cut(0)
c2 = dummy_cut(0)
c2.id = "dummy-cut-0001" # override cut ID, retain identical recording ID as `c`
cuts = CutSet.from_cuts([c, c2])
recs, sups, feats = cuts.decompose()
assert isinstance(recs, RecordingSet)
# deduplicated recording
assert len(recs) == 1
assert recs[0].id == "dummy-recording-0000"
assert sups is None
assert isinstance(feats, FeatureSet)
# not deduplicated features
assert len(feats) == 2
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 test_mono_cut_fill_supervision_shrink():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut.duration = 0.5
fcut = cut.fill_supervision(shrink_ok=True)
# Original is not modified
assert cut.supervisions[0].start == 0
assert cut.supervisions[0].duration == 1
# Result is modified
assert fcut.supervisions[0].start == 0
assert fcut.supervisions[0].duration == 0.5
def test_mixed_cut_fill_supervision_expand():
cut = dummy_cut(0, supervisions=[dummy_supervision(0)])
cut = cut.pad(duration=7.51)
fcut = cut.fill_supervision()
# Original is not modified
assert cut.supervisions[0].start == 0
assert cut.supervisions[0].duration == 1
# Result is modified
assert fcut.supervisions[0].start == 0
assert fcut.supervisions[0].duration == 7.51
def test_single_cut_sampler_len():
# total duration is 55 seconds
# each second has 100 frames
cuts = CutSet.from_cuts(dummy_cut(idx, duration=float(idx)) for idx in range(1, 11))
sampler = SimpleCutSampler(cuts, shuffle=True, max_frames=10 * 100, max_cuts=6)
for epoch in range(5):
sampler.set_epoch(epoch)
sampler_len = len(sampler)
num_batches = len([batch for batch in sampler])
assert sampler_len == num_batches
def test_extend_by_cut(
cut_start,
cut_duration,
extend_duration,
extend_direction,
expected_start,
expected_end,
):
cut = dummy_cut(int(uuid4()), start=cut_start, duration=cut_duration)
extended_cut = cut.extend_by(duration=extend_duration, direction=extend_direction)
assert isclose(extended_cut.start, expected_start)
assert isclose(extended_cut.end, expected_end)
def test_cut_load_array_truncate():
"""Check that loading a custom Array works after truncation."""
ivector = np.arange(20).astype(np.float32)
with NamedTemporaryFile(suffix=".h5") as f, LilcomHdf5Writer(
f.name) as writer:
cut = dummy_cut(0, duration=5.0)
cut.ivector = writer.store_array(key="utt1", value=ivector)
cut = cut.truncate(duration=3)
restored_ivector = cut.load_ivector()
np.testing.assert_equal(ivector, restored_ivector)
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 test_cut_pairs_sampler_len():
# total duration is 55 seconds
# each second has 100 frames
cuts = CutSet.from_cuts(dummy_cut(idx, duration=float(idx)) for idx in range(1, 11))
sampler = CutPairsSampler(
source_cuts=cuts,
target_cuts=cuts,
shuffle=True,
max_source_frames=10 * 100,
max_target_frames=10 * 100,
)
for epoch in range(5):
assert len(sampler) == len([batch for batch in sampler])
sampler.set_epoch(epoch)
def test_bucketing_sampler_len():
# total duration is 550 seconds
# each second has 100 frames
cuts = CutSet.from_cuts(
dummy_cut(idx, duration=float(duration))
for idx, duration in enumerate(list(range(1, 11)) * 10))
sampler = BucketingSampler(cuts,
num_buckets=4,
shuffle=True,
max_frames=64 * 100,
max_cuts=6)
for epoch in range(5):
assert len(sampler) == len([item for item in sampler])
sampler.set_epoch(epoch)
def test_cut_load_temporal_array_truncate():
"""Check the array loaded via TemporalArray is truncated along with the cut."""
with NamedTemporaryFile(suffix=".h5") as f, NumpyHdf5Writer(
f.name) as writer:
expected_duration = 52.4 # 131 frames x 0.4s frame shift == 52.4s
cut = dummy_cut(0, duration=expected_duration)
alignment = np.random.randint(500, size=131)
cut.alignment = writer.store_array(key="utt1",
value=alignment,
frame_shift=0.4,
temporal_dim=0)
cut_trunc = cut.truncate(duration=5.0)
alignment_piece = cut_trunc.load_alignment()
assert alignment_piece.shape == (13, ) # 5.0 / 0.4 == 12.5 ~= 13
np.testing.assert_equal(alignment[:13], alignment_piece)
def test_concat_cuts():
cuts = [
dummy_cut(duration=30.0),
dummy_cut(duration=20.0),
dummy_cut(duration=10.0),
dummy_cut(duration=5.0),
dummy_cut(duration=4.0),
dummy_cut(duration=3.0),
dummy_cut(duration=2.0),
]
concat = concat_cuts(cuts, gap=1.0)
assert [c.duration for c in concat] == [
30.0,
20.0 + 1.0 + 2.0 + 1.0 + 3.0, # == 27.0
10.0 + 1.0 + 4.0 + 1.0 + 5.0, # == 21.0
]