def test_sine_mfcc_time_major(self):
sine = utils.make_test_sine(2048, 440, 44100)
audio = AudioBuffer(sine, 44100)
mfcc = MFCC(hop_size=512, frame_size=1024, time_major=True)
features = mfcc(audio)
assert features.shape == (5, 20)
def test_sine_wave_fft_power(self):
bin_freq = 44100. / 1024.
sine = utils.make_test_sine(1024, bin_freq * 10, 44100)
audio = AudioBuffer(sine, 44100)
fft = FFT(output='power')
features = fft(audio)
expected = np.zeros(513)
expected[10] = 512. * 512.
assert features.shape == (513, )
assert isinstance(features[0], np.float32)
np.testing.assert_array_almost_equal(features, expected)
def test_sine_wave_fft_magnitude_longer(self):
bin_freq = 44100. / 4096.
sine = utils.make_test_sine(4096, bin_freq * 10, 44100)
audio = AudioBuffer(sine, 44100)
fft = FFT(output='magnitude')
features = fft(audio)
expected = np.zeros(2049)
expected[10] = 2048.
assert features.shape == (2049, )
assert isinstance(features[0], np.float32)
np.testing.assert_array_almost_equal(features, expected)
def test_sine_mfcc_scale(self):
batch_size = 10
mfcc = MFCC(hop_size=512, frame_size=1024, scale_axis=(0, 2))
features = np.zeros((batch_size, 20, 5))
for i in range(batch_size):
sine = utils.make_test_sine(2048, 100 + (50 * i), 44100)
audio = AudioBuffer(sine, 44100)
features[i] = mfcc(audio)
assert features.shape == (10, 20, 5)
scaled = mfcc.fit_scaler(features)
assert scaled.mean() == pytest.approx(0.)
assert scaled.std() == pytest.approx(1.)
np.testing.assert_array_almost_equal(scaled.mean((0, 2)), np.zeros(20))
np.testing.assert_array_almost_equal(scaled.std((0, 2)), np.ones(20))