def test_inverse_stft_window_fn(self):
"""Test that inverse_stft_window_fn has unit gain at each window phase."""
# Tuples of (frame_length, frame_step).
test_configs = [
(256, 32),
(256, 64),
(128, 25),
(127, 32),
(128, 64),
]
for (frame_length, frame_step) in test_configs:
hann_window = window_ops.hann_window(frame_length,
dtype=dtypes.float32)
inverse_window_fn = spectral_ops.inverse_stft_window_fn(frame_step)
inverse_window = inverse_window_fn(frame_length,
dtype=dtypes.float32)
with self.cached_session(use_gpu=True) as sess:
hann_window, inverse_window = self.evaluate(
[hann_window, inverse_window])
# Expect unit gain at each phase of the window.
product_window = hann_window * inverse_window
for i in range(frame_step):
self.assertAllClose(1.0, np.sum(product_window[i::frame_step]))
def test_inverse_stft_window_fn_special_case(self):
"""Test inverse_stft_window_fn in special overlap = 3/4 case."""
# Cases in which frame_length is an integer multiple of 4 * frame_step are
# special because they allow exact reproduction of the waveform with a
# squared Hann window (Hann window in both forward and reverse transforms).
# In the case where frame_length = 4 * frame_step, that combination
# produces a constant gain of 1.5, and so the corrected window will be the
# Hann window / 1.5.
# Tuples of (frame_length, frame_step).
test_configs = [
(256, 64),
(128, 32),
]
for (frame_length, frame_step) in test_configs:
hann_window = window_ops.hann_window(frame_length,
dtype=dtypes.float32)
inverse_window_fn = spectral_ops.inverse_stft_window_fn(frame_step)
inverse_window = inverse_window_fn(frame_length,
dtype=dtypes.float32)
with self.cached_session(use_gpu=True) as sess:
hann_window, inverse_window = self.evaluate(
[hann_window, inverse_window])
self.assertAllClose(hann_window, inverse_window * 1.5)
def test_inverse_stft_window_fn_special_case(self):
"""Test inverse_stft_window_fn in special overlap = 3/4 case."""
# Cases in which frame_length is an integer multiple of 4 * frame_step are
# special because they allow exact reproduction of the waveform with a
# squared Hann window (Hann window in both forward and reverse transforms).
# In the case where frame_length = 4 * frame_step, that combination
# produces a constant gain of 1.5, and so the corrected window will be the
# Hann window / 1.5.
# Tuples of (frame_length, frame_step).
test_configs = [
(256, 64),
(128, 32),
]
for (frame_length, frame_step) in test_configs:
hann_window = window_ops.hann_window(frame_length, dtype=dtypes.float32)
inverse_window_fn = spectral_ops.inverse_stft_window_fn(frame_step)
inverse_window = inverse_window_fn(frame_length, dtype=dtypes.float32)
with self.cached_session(use_gpu=True) as sess:
hann_window, inverse_window = self.evaluate(
[hann_window, inverse_window])
self.assertAllClose(hann_window, inverse_window * 1.5)
def compute_spectrogram_tf(
waveform,
frame_length=2048, frame_step=512,
spec_exponent=1., window_exponent=1.):
""" Compute magnitude / power spectrogram from waveform as
a n_samples x n_channels tensor.
:param waveform: Input waveform as (times x number of channels)
tensor.
:param frame_length: Length of a STFT frame to use.
:param frame_step: HOP between successive frames.
:param spec_exponent: Exponent of the spectrogram (usually 1 for
magnitude spectrogram, or 2 for power spectrogram).
:param window_exponent: Exponent applied to the Hann windowing function
(may be useful for making perfect STFT/iSTFT
reconstruction).
:returns: Computed magnitude / power spectrogram as a
(T x F x n_channels) tensor.
"""
stft_tensor = tf.transpose(
stft(
tf.transpose(waveform),
frame_length,
frame_step,
window_fn=lambda f, dtype: hann_window(
f,
periodic=True,
dtype=waveform.dtype) ** window_exponent),
perm=[1, 2, 0])
return tf.abs(stft_tensor) ** spec_exponent
def test_inverse_stft_window_fn(self, frame_length, frame_step):
"""Test that inverse_stft_window_fn has unit gain at each window phase."""
hann_window = window_ops.hann_window(frame_length, dtype=dtypes.float32)
inverse_window_fn = spectral_ops.inverse_stft_window_fn(frame_step)
inverse_window = inverse_window_fn(frame_length, dtype=dtypes.float32)
hann_window, inverse_window = self.evaluate([hann_window, inverse_window])
# Expect unit gain at each phase of the window.
product_window = hann_window * inverse_window
for i in range(frame_step):
self.assertAllClose(1.0, np.sum(product_window[i::frame_step]))
def test_inverse_stft_window_fn_special_case(self, frame_length, frame_step): """Test inverse_stft_window_fn in special overlap = 3/4 case.""" # Cases in which frame_length is an integer multiple of 4 * frame_step are # special because they allow exact reproduction of the waveform with a # squared Hann window (Hann window in both forward and reverse transforms). # In the case where frame_length = 4 * frame_step, that combination # produces a constant gain of 1.5, and so the corrected window will be the # Hann window / 1.5. hann_window = window_ops.hann_window(frame_length, dtype=dtypes.float32) inverse_window_fn = spectral_ops.inverse_stft_window_fn(frame_step) inverse_window = inverse_window_fn(frame_length, dtype=dtypes.float32) self.assertAllClose(hann_window, inverse_window * 1.5)
def test_inverse_stft_window_fn(self):
"""Test that inverse_stft_window_fn has unit gain at each window phase."""
# Tuples of (frame_length, frame_step).
test_configs = [
(256, 32),
(256, 64),
(128, 25),
(127, 32),
(128, 64),
]
for (frame_length, frame_step) in test_configs:
hann_window = window_ops.hann_window(frame_length, dtype=dtypes.float32)
inverse_window_fn = spectral_ops.inverse_stft_window_fn(frame_step)
inverse_window = inverse_window_fn(frame_length, dtype=dtypes.float32)
with self.cached_session(use_gpu=True) as sess:
hann_window, inverse_window = self.evaluate(
[hann_window, inverse_window])
# Expect unit gain at each phase of the window.
product_window = hann_window * inverse_window
for i in range(frame_step):
self.assertAllClose(1.0, np.sum(product_window[i::frame_step]))