def __call__(self, original, sample_rate):
"""Augment original waveform
Parameters
----------
original : `np.ndarray`
(n_samples, n_channels) waveform.
sample_rate : `int`
Sample rate.
Returns
-------
augmented : `np.ndarray`
(n_samples, n_channels) noise-augmented waveform.
"""
raw_audio = RawAudio(sample_rate=sample_rate, mono=True)
# accumulate enough noise to cover duration of original waveform
noises = []
len_left = len(original)
while len_left > 0:
# select noise file at random
file = random.choice(self.files_)
# select noise segment at random
segment = next(random_segment(file['gaps'], weighted=False))
duration = segment.duration
segment_len = duration * sample_rate
# if noise segment is longer than what is needed, crop it at random
if segment_len > len_left:
duration = len_left / sample_rate
segment = next(random_subsegment(segment, duration))
noise = raw_audio.crop(file,
segment,
mode='center',
fixed=duration)
# decrease the `len_left` value by the size of the returned noise
len_left -= len(noise)
noise = normalize(noise)
noises.append(noise)
# concatenate
# FIXME: use fade-in between concatenated noises
noise = np.vstack(noises)
# select SNR at random
snr = (self.snr_max -
self.snr_min) * np.random.random_sample() + self.snr_min
alpha = np.exp(-np.log(10) * snr / 20)
return normalize(original) + alpha * noise
def __call__(self, original, sample_rate):
"""Augment original waveform
Parameters
----------
original : `np.ndarray`
(n_samples, n_channels) waveform.
sample_rate : `int`
Sample rate.
Returns
-------
augmented : `np.ndarray`
(n_samples, n_channels) noise-augmented waveform.
"""
raw_audio = RawAudio(sample_rate=sample_rate, mono=True)
original_duration = len(original) / sample_rate
# accumulate enough noise to cover duration of original waveform
noises = []
left = original_duration
while left > 0:
# select noise file at random
file = random.choice(self.files_)
duration = file['duration']
# if noise file is longer than what is needed, crop it
if duration > left:
segment = next(random_subsegment(Segment(0, duration), left))
noise = raw_audio.crop(file,
segment,
mode='center',
fixed=left)
left = 0
# otherwise, take the whole file
else:
noise = raw_audio(file).data
left -= duration
noise = normalize(noise)
noises.append(noise)
# concatenate
# FIXME: use fade-in between concatenated noises
noise = np.vstack(noises)
# select SNR at random
snr = (self.snr_max -
self.snr_min) * np.random.random_sample() + self.snr_min
alpha = np.exp(-np.log(10) * snr / 20)
return normalize(original) + alpha * noise
def _create_config(self, segment_size_sec: float):
return metrics.SpeakerValidationConfig(
protocol_name='VoxCeleb.SpeakerVerification.VoxCeleb2',
feature_extraction=RawAudio(sample_rate=self.sample_rate),
preprocessors={'audio': FileFinder()},
duration=segment_size_sec)