def train_on_audio(self, fn: str):
"""Run through a single audio file"""
save_test = random() > 0.8
audio = load_audio(fn)
num_chunks = len(audio) // self.args.chunk_size
self.listener.clear()
for i, chunk in enumerate(chunk_audio(audio, self.args.chunk_size)):
print('\r' + str(i * 100. / num_chunks) + '%', end='', flush=True)
self.audio_buffer = np.concatenate((self.audio_buffer[len(chunk):], chunk))
conf = self.listener.update(chunk)
if conf > self.args.threshold:
self.samples_since_train += 1
name = splitext(basename(fn))[0] + '-' + str(i) + '.wav'
name = join(self.args.folder, 'test' if save_test else '', 'not-wake-word',
'generated', name)
save_audio(name, self.audio_buffer)
print()
print('Saved to:', name)
if not save_test and self.samples_since_train >= self.args.delay_samples and \
self.args.epochs > 0:
self.samples_since_train = 0
self.retrain()
def chunk_audio_pieces(self, pieces, chunk_size):
"""Convert chunks of audio into a series of equally sized pieces"""
left_over = np.array([])
for piece in pieces:
if left_over.size == 0:
combined = piece
else:
combined = np.concatenate([left_over, piece], axis=-1)
for chunk in chunk_audio(combined.T, chunk_size):
yield chunk.T
left_over = piece[-(len(piece) % chunk_size):]
def vectors_from_fn(self, fn: str):
"""
Run through a single background audio file, overlaying with wake words.
Generates (mfccs, target) where mfccs is a series of mfcc values and
target is a single integer classification of the target network output for that chunk
"""
audio = load_audio(fn)
audio_volume = self.calc_volume(audio)
audio_volume *= 0.4 + 0.5 * random()
audio = self.normalize_volume_to(audio, audio_volume)
self.listener.clear()
chunked_bg = chunk_audio(audio, self.args.chunk_size)
chunked_ww = self.chunk_audio_pieces(
self.generate_wakeword_pieces(audio_volume), self.args.chunk_size)
for i, (chunk_bg, (chunk_ww,
targets)) in enumerate(zip(chunked_bg, chunked_ww)):
chunk = self.merge(chunk_bg, chunk_ww, 0.6)
self.vals_buffer = np.concatenate(
(self.vals_buffer[len(targets):], targets))
self.audio_buffer = np.concatenate(
(self.audio_buffer[len(chunk):], chunk))
mfccs = self.listener.update_vectors(chunk)
percent_overlapping = self.max_run_length(
self.vals_buffer, 1) / len(self.vals_buffer)
if self.vals_buffer[-1] == 0 and percent_overlapping > 0.8:
target = 1
elif percent_overlapping < 0.5:
target = 0
else:
continue
if random() > 1.0 - self.args.save_prob:
name = splitext(basename(fn))[0]
wav_file = join('debug', 'ww' if target == 1 else 'nww',
'{} - {}.wav'.format(name, i))
save_audio(wav_file, self.audio_buffer)
yield mfccs, target