yuzu/spleeter
1
0
Fork 0
mirror of https://github.com/YuzuZensai/spleeter.git synced 2026-01-06 04:32:43 +00:00
Code Issues Packages Projects Releases Wiki Activity

dont pad the signal, only the stft matrix before inversion

Browse Source
This commit is contained in:
mmoussallam
2020-06-19 00:33:19 +02:00
parent 94dedbe949
commit 97458816c9
2 changed files with 13 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
spleeter/separator.py
View File

@@ -122,20 +122,21 @@ class Separator(object):
assert not (inverse and length is None)
data = np.asfortranarray(data)
N = self._params["frame_length"]
pad_edges = int(N/2)
H = self._params["frame_step"]
F = int(N/2) + 1
win = hann(N, sym=False)
fstft = istft if inverse else stft
win_len_arg = {"win_length": None, "length": length +
2*pad_edges} if inverse else {"n_fft": N}
win_len_arg = {"win_length": None,
"length": None} if inverse else {"n_fft": N}
n_channels = data.shape[-1]
out = []
for c in range(n_channels):
d = data[:, :, c].T if inverse else np.concatenate(
(np.zeros(pad_edges,), data[:, c], np.zeros(pad_edges,)))
d = np.concatenate((np.zeros((F, 1)), data[:, :, c].T, np.zeros(
(F, 1))), axis=1) if inverse else data[:, c]
s = fstft(d, hop_length=H, window=win, center=False, **win_len_arg)
if inverse:
s = s[pad_edges:-pad_edges]
s = s[H:]
s = s[:length]
s = np.expand_dims(s.T, 2-inverse)
out.append(s)
if len(out) == 1:

12
tests/test_separator.py
View File

@@ -51,7 +51,7 @@ def test_separator_backends(test_file):
stft_matrix = separator_lib._stft(waveform)
reconstructed = separator_lib._stft(
stft_matrix, inverse=True, length=waveform.shape[0])
assert np.allclose(reconstructed, waveform, atol=1e-2)
assert np.allclose(reconstructed, waveform, atol=3e-2)
# # now also test that tensorflow and librosa STFT provide same results
from spleeter.audio.spectrogram import compute_spectrogram_tf
@@ -62,11 +62,10 @@ def test_separator_backends(test_file):
with tf.Session() as sess:
spectrogram_tf_eval = spectrogram_tf.eval()
# check that stfts are equivalent up to the padding in the librosa case
assert stft_matrix.shape[0] == spectrogram_tf_eval.shape[0] + 2
assert stft_matrix.shape[1:] == spectrogram_tf_eval.shape[1:]
# check that stfts are equivalent
assert stft_matrix.shape == spectrogram_tf_eval.shape
assert np.allclose(
np.abs(stft_matrix[1:-1]), spectrogram_tf_eval, atol=1e-2)
np.abs(stft_matrix), spectrogram_tf_eval, atol=1e-2)
# compare both separation, it should be close
out_tf = separator_tf._separate_tensorflow(waveform, test_file)
@@ -78,7 +77,8 @@ def test_separator_backends(test_file):
print(np.sum(np.abs(out_lib[instrument])))
assert np.sum(np.abs(out_tf[instrument])) > 1000
assert np.sum(np.abs(out_lib[instrument])) > 1000
assert np.allclose(out_tf[instrument], out_lib[instrument], atol=0.01)
print(np.max(out_tf[instrument]- out_lib[instrument]))
assert np.allclose(out_tf[instrument], out_lib[instrument], atol=0.025)
@pytest.mark.parametrize('test_file, configuration, backend', TEST_CONFIGURATIONS)