def stft_to_feature(self, stft_signals):
if pt_complex.is_torch(stft_signals):
concatenate = torch.cat
else:
concatenate = np.concatenate
return concatenate([
abs(stft_signals),
stft_signals.real,
stft_signals.imag,
],
axis=-1)
def stft_to_feature(self, stft_signals):
if pt_complex.is_torch(stft_signals):
raise NotImplementedError()
return torch.log1p(abs(stft_signals))
else:
angle = np.angle(stft_signals)
return np.concatenate([
np.log1p(abs(stft_signals)),
np.cos(angle),
np.sin(angle),
],
axis=-1)
def stft(self, signal):
kwargs = dict(
size=self.size,
shift=self.shift,
axis=-1,
window_length=self.window_length,
window=self._window,
fading=self.fading,
pad=self.pad,
# symmetric_window=False,
)
if pt_complex.is_torch(signal):
return pt_stft(signal, **kwargs)
else:
return pb.transform.stft(signal, **kwargs)
def stft_to_feature(self, stft_signals):
if pt_complex.is_torch(stft_signals):
return torch.log1p(abs(stft_signals))
else:
return np.log1p(abs(stft_signals))
def istft(self, signal, num_samples=None):
"""
Args:
signal:
samples:
Returns:
>>> def print_properties(array):
... print(f'array(shape={array.shape}, dtype={array.dtype})')
>>> fe_stft = STFT()
>>> fe_stft
STFT(size=1024, shift=256, window_length=1024, pad=True, fading=True, output_size=513, window='blackman')
>>> audio = pb.io.load('/net/vol/jenkins/python_unittest_data/timit/data/sample_1.wav')
>>> print_properties(audio)
array(shape=(46797,), dtype=float64)
>>> samples, = audio.shape
>>> print_properties(fe_stft([audio]))
array(shape=(1, 186, 513), dtype=complex128)
>>> print_properties(fe_stft.istft(fe_stft([audio])))
array(shape=(1, 46848), dtype=float64)
>>> print_properties(fe_stft.istft(fe_stft([audio]), num_samples=samples))
array(shape=(1, 46797), dtype=float64)
>>> print_properties(fe_stft.istft(fe_stft(np.zeros([1023]))))
array(shape=(1024,), dtype=float64)
>>> print_properties(fe_stft.istft(fe_stft(np.zeros([1024]))))
array(shape=(1024,), dtype=float64)
>>> print_properties(fe_stft.istft(fe_stft(np.zeros([1025]))))
array(shape=(1280,), dtype=float64)
>>> print_properties(fe_stft.istft(fe_stft(np.zeros([1023])), num_samples=1023))
array(shape=(1023,), dtype=float64)
>>> print_properties(fe_stft.istft(fe_stft(np.zeros([1024])), num_samples=1024))
array(shape=(1024,), dtype=float64)
>>> print_properties(fe_stft.istft(fe_stft(np.zeros([1025])), num_samples=1025))
array(shape=(1025,), dtype=float64)
"""
kwargs = dict(
size=self.size,
shift=self.shift,
# axis=axis,
window=self._window,
window_length=self.window_length,
fading=self.fading,
pad=self.pad,
# symmetric_window=False,
# dft_norm=self.dft_norm,
num_samples=num_samples,
)
if pt_complex.is_torch(signal):
time_signal = pt_istft(signal, **kwargs)
else:
time_signal = pb.transform.istft(signal, **kwargs)
return time_signal