def main():
args = create_parser(usage).parse_args()
args.tags_file = abspath(args.tags_file) if args.tags_file else None
args.folder = abspath(args.folder)
args.output_folder = abspath(args.output_folder)
noise_min, noise_max = args.noise_ratio_low, args.noise_ratio_high
data = TrainData.from_both(args.tags_file, args.folder, args.folder)
noise_data = NoiseData(args.noise_folder)
print('Data:', data)
def translate_filename(source: str, n=0) -> str:
assert source.startswith(args.folder)
relative_file = source[len(args.folder):].strip(os.path.sep)
if n > 0:
base, ext = splitext(relative_file)
relative_file = base + '.' + str(n) + ext
return join(args.output_folder, relative_file)
all_filenames = sum(data.train_files + data.test_files, [])
for i, filename in enumerate(all_filenames):
print('{0:.2%} \r'.format(i / (len(all_filenames) - 1)), end='', flush=True)
audio = load_audio(filename)
for n in range(args.inflation_factor):
altered = noise_data.noised_audio(audio, noise_min + (noise_max - noise_min) * random())
output_filename = translate_filename(filename, n)
makedirs(dirname(output_filename), exist_ok=True)
save_audio(output_filename, altered)
print('Done!')
if args.tags_file and args.tags_file.startswith(args.folder):
shutil.copy2(args.tags_file, translate_filename(args.tags_file))
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 on_activation_safe():
global chunk_num
nm = join(
args.save_dir,
args.save_prefix + session_id + '.' + str(chunk_num) + '.wav')
save_audio(nm, audio_buffer)
print()
print('Saved to ' + nm + '.')
chunk_num += 1
def on_activation(self):
activate_notify()
if self.args.save_dir:
nm = join(
self.args.save_dir, self.args.save_prefix + self.session_id +
'.' + str(self.chunk_num) + '.wav')
save_audio(nm, self.audio_buffer)
print()
print('Saved to ' + nm + '.')
self.chunk_num += 1
def on_activation():
Popen(['aplay', '-q', 'data/activate.wav'])
if args.save_dir:
global chunk_num
nm = join(
args.save_dir,
args.save_prefix + session_id + '.' + str(chunk_num) + '.wav')
save_audio(nm, audio_buffer)
print()
print('Saved to ' + nm + '.')
chunk_num += 1
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
def generate_samples(self, folder, name, value, duration):
for i in range(self.count):
save_audio(join(folder, name.format(i)),
np.array([value] * int(duration * pr.sample_rate)))
