def generate(hop_size=256):
while True:
shuffled = sklearn.utils.shuffle(files)
for f in shuffled:
audio, _ = malaya_speech.load(f, sr=sr)
mel = malaya_speech.featurization.universal_mel(audio)
batch_max_steps = random.randint(16384, 110_250)
batch_max_frames = batch_max_steps // hop_size
if len(mel) > batch_max_frames:
interval_start = 0
interval_end = len(mel) - batch_max_frames
start_frame = random.randint(interval_start, interval_end)
start_step = start_frame * hop_size
audio = audio[start_step:start_step + batch_max_steps]
mel = mel[start_frame:start_frame + batch_max_frames, :]
v = speaker_model([audio])
yield {
'mel': mel,
'mel_length': [len(mel)],
'audio': audio,
'v': v[0],
}
def signal_augmentation(wav):
seed = random.randint(0, 100_000_000)
wav = calc(wav, seed)
if random.gauss(0.5, 0.14) > 0.6:
n, _ = malaya_speech.load(random.choice(noises), sr = 16000)
n = calc(n, seed, True)
combined = augmentation.add_noise(
wav, n, factor = random.uniform(0.05, 0.3)
)
else:
combined = wav
return combined.astype('float32')
def read_wav(f):
return malaya_speech.load(f, sr = sr)
