def test_time_diff_to_num_frames(time_diff: Seconds, frame_length: Seconds,
frame_shift: Seconds,
expected_num_frames: int):
assert time_diff_to_num_frames(
time_diff=time_diff,
frame_length=frame_length,
frame_shift=frame_shift) == expected_num_frames
def test_vad_dataset(cut_set):
duration = 5.0
feature_threhold = -8.0
cuts = list(cut_set)
assert isclose(cuts[0].duration, 6.0)
assert len(cuts[0].supervisions) == 2
v1_start = time_diff_to_num_frames(cuts[0].supervisions[0].start,
frame_length=0,
frame_shift=0.01)
v1_end = time_diff_to_num_frames(cuts[0].supervisions[0].end,
frame_length=0,
frame_shift=0.01)
v2_start = time_diff_to_num_frames(cuts[0].supervisions[1].start,
frame_length=0,
frame_shift=0.01)
v2_end = time_diff_to_num_frames(cuts[0].supervisions[1].end,
frame_length=0,
frame_shift=0.01)
end = time_diff_to_num_frames(duration, frame_length=0, frame_shift=0.01)
# Convert long cuts into 5s cuts
window_cut_set = cut_set.cut_into_windows(duration=duration)
# Create a one-element list with the ID of the first cut
first_id = [next(iter(window_cut_set.ids))]
dataset = VadDataset(window_cut_set)
example = dataset[first_id]
is_voice = example['is_voice'][0]
assert isclose(float(torch.mean(is_voice[0:v1_start])), 0)
assert isclose(float(torch.mean(is_voice[v1_start:v1_end])), 1)
assert isclose(float(torch.mean(is_voice[v1_end:v2_start])), 0)
assert isclose(float(torch.mean(is_voice[v2_start:v2_end])), 1)
assert isclose(float(torch.mean(is_voice[v2_end:end])), 0)
features = example['inputs'][0]
assert float(torch.mean(features[0:v1_start])) < feature_threhold
assert float(torch.mean(features[v1_start:v1_end])) > feature_threhold
assert float(torch.mean(features[v1_end:v2_start])) < feature_threhold
assert float(torch.mean(features[v2_start:v2_end])) > feature_threhold
assert float(torch.mean(features[v2_end:end])) < feature_threhold
cut = example['cut'][0]
assert cut.duration == duration
def test_vad_dataset(cut_set):
duration = 5.0
feature_threhold = -8.0
cuts = list(cut_set)
assert isclose(cuts[0].duration, 6.0)
assert len(cuts[0].supervisions) == 2
v1_start = time_diff_to_num_frames(cuts[0].supervisions[0].start,
frame_length=0,
frame_shift=0.01)
v1_end = time_diff_to_num_frames(cuts[0].supervisions[0].end,
frame_length=0,
frame_shift=0.01)
v2_start = time_diff_to_num_frames(cuts[0].supervisions[1].start,
frame_length=0,
frame_shift=0.01)
v2_end = time_diff_to_num_frames(cuts[0].supervisions[1].end,
frame_length=0,
frame_shift=0.01)
end = time_diff_to_num_frames(duration, frame_length=0, frame_shift=0.01)
dataset = VadDataset(cut_set.cut_into_windows(duration=duration))
example = dataset[0]
assert isclose(float(torch.mean(example['is_voice'][0:v1_start])), 0)
assert isclose(float(torch.mean(example['is_voice'][v1_start:v1_end])), 1)
assert isclose(float(torch.mean(example['is_voice'][v1_end:v2_start])), 0)
assert isclose(float(torch.mean(example['is_voice'][v2_start:v2_end])), 1)
assert isclose(float(torch.mean(example['is_voice'][v2_end:end])), 0)
assert float(torch.mean(
example['features'][0:v1_start])) < feature_threhold
assert float(torch.mean(
example['features'][v1_start:v1_end])) > feature_threhold
assert float(torch.mean(
example['features'][v1_end:v2_start])) < feature_threhold
assert float(torch.mean(
example['features'][v2_start:v2_end])) > feature_threhold
assert float(torch.mean(
example['features'][v2_end:end])) < feature_threhold
