def test_kaldifeat_supports_list_of_uneven_len_inputs():
input = [
torch.arange(8000, dtype=torch.float32),
torch.arange(16000, dtype=torch.float32),
]
fe = KaldifeatFbank()
feats = fe.extract(input, sampling_rate=16000)
assert len(feats) == 2
f1, f2 = feats
assert f1.shape == (50, 80)
assert f2.shape == (100, 80)
def test_kaldifeat_supports_list_of_even_len_inputs(input):
fe = KaldifeatFbank()
feats = fe.extract(input, sampling_rate=16000)
assert feats.ndim == 3
assert feats.shape == (2, 50, 80)
def test_kaldifeat_supports_list_with_single_input_waveform(input):
fe = KaldifeatFbank()
feats = fe.extract(input, sampling_rate=16000)
assert isinstance(feats, list)
assert len(feats) == 1
assert feats[0].shape == (50, 80)
def test_kaldifeat_supports_single_input_waveform(input):
fe = KaldifeatFbank()
feats = fe.extract(input, sampling_rate=16000)
assert feats.shape == (50, 80)
def test_kaldifeat_config():
x = np.arange(8000, dtype=np.float32)
fe = KaldifeatFbank(KaldifeatFbankConfig(mel_opts=KaldifeatMelOptions(num_bins=27)))
feats = fe.extract(x, sampling_rate=16000)
assert feats.shape == (50, 27)
input = [
torch.arange(8000, dtype=torch.float32),
torch.arange(16000, dtype=torch.float32),
]
fe = KaldifeatFbank()
feats = fe.extract(input, sampling_rate=16000)
assert len(feats) == 2
f1, f2 = feats
assert f1.shape == (50, 80)
assert f2.shape == (100, 80)
@pytest.mark.parametrize(
["extractor1", "extractor2"],
[
(KaldifeatFbank(), Fbank()),
(KaldifeatMfcc(), Mfcc()),
],
)
def test_kaldifeat_torchaudio_equivalence(extractor1, extractor2):
sampling_rate = 16000
np.random.seed(99) # ensure reproducibility
audio = np.random.rand(1, 32000).astype(np.float32)
feat1 = extractor1.extract(audio, sampling_rate)
feat2 = extractor2.extract(audio, sampling_rate)
np.testing.assert_almost_equal(feat1, feat2, decimal=3)
@pytest.mark.parametrize("feature_type", ["kaldifeat-fbank", "kaldifeat-mfcc"])
def test_feature_extractor_serialization(feature_type):
fe = create_default_feature_extractor(feature_type)