def test_frequency_filter_gives_correct_size(self, n_frequencies, n_frames,
window_size):
"""Tests filtering signals with frequency sampling method.
Generate random signals and magnitudes and filter using fft_convolve().
Args:
n_frequencies: Number of magnitudes.
n_frames: Number of frames for a time-varying filter.
window_size: Size of window for generating impulse responses.
"""
# Create transfer function.
if n_frames > 0:
magnitudes = np.random.uniform(size=(1, n_frames,
n_frequencies)).astype(
np.float32)
else:
magnitudes = np.random.uniform(size=(1, n_frequencies)).astype(
np.float32)
audio_out = core.frequency_filter(self.audio,
magnitudes,
window_size=window_size,
padding='same')
audio_out_size = int(audio_out.shape[-1])
self.assertEqual(audio_out_size, self.audio_size)
def get_signal(self, audio, magnitudes):
"""Filter audio with LTV-FIR filter.
Args:
audio: Dry audio. 2-D Tensor of shape [batch, n_samples].
magnitudes: Magnitudes tensor of shape [batch, n_frames, n_filter_banks].
Expects float32 that is strictly positive.
Returns:
signal: Filtered audio of shape [batch, n_samples, 1].
"""
return core.frequency_filter(audio,
magnitudes,
window_size=self.window_size)
def get_signal(self, magnitudes):
"""Synthesize audio with filtered white noise.
Args:
magnitudes: Magnitudes tensor of shape [batch, n_frames, n_filter_banks].
Expects float32 that is strictly positive.
Returns:
signal: A tensor of harmonic waves of shape [batch, n_samples, 1].
"""
batch_size = int(magnitudes.shape[0])
signal = tf.random.uniform(
[batch_size, self.n_samples], minval=-1.0, maxval=1.0)
return core.frequency_filter(signal,
magnitudes,
window_size=self.window_size)
