def test_sine_wave_stft_complex_time_major(self, shared_datadir):
file_name = (shared_datadir / 'audio/test_sine.wav').resolve()
audio = AudioBuffer()
audio.load(file_name)
stft = STFT(time_major=True, fft_size=2048, hop_size=1024)
features = stft(audio)
assert features.shape == (44, 1025)
assert isinstance(features[0][0], np.complex64)
def test_sine_wave_stft_complex(self, shared_datadir):
file_name = (shared_datadir / 'audio/test_sine.wav').resolve()
audio = AudioBuffer()
audio.load(file_name)
stft = STFT()
features = stft(audio)
assert features.shape == (513, 87)
assert isinstance(features[0][0], np.complex64)
def test_sine_wave_stft_magnitude_scaled(self):
bin_freq = 44100. / 1024.
length = (1024. * 10) + 1
stft = STFT(output='magnitude',
fft_size=1024,
hop_size=512,
time_major=True,
scale_axis=None)
batch_size = 10
features = np.zeros((10, 21, 513))
for i in range(batch_size):
sine = utils.make_test_cosine(int(length), bin_freq * (i + 10),
44100)
audio = AudioBuffer(sine, 44100)
features[i] = stft(audio)
scaled = stft.fit_scaler(features)
assert scaled.shape == (10, 21, 513)
assert scaled.mean() == pytest.approx(0.)
assert scaled.std() == pytest.approx(1.)
def test_sine_wave_stft_power(self):
bin_freq = 44100. / 1024.
sine = utils.make_test_cosine(1024 + 1, bin_freq * 10, 44100)
audio = AudioBuffer(sine, 44100)
stft = STFT(output='power', fft_size=1024)
features = stft(audio)
expected = np.zeros((513, 3))
# amplitude is spread out over neighbouring bins
expected[9, :] = 128. * 128.
expected[10, :] = 256. * 256.
expected[11, :] = 128. * 128.
assert features.shape == (513, 3)
assert isinstance(features[0][0], np.float32)
np.testing.assert_array_almost_equal(features, expected)
def test_empty_construction(self):
stft = STFT()