def _process_utterance(in_path, out_path, speaker, text):
# Change sampling rate
try:
old_samplerate, old_audio = read_wav_np(in_path)
except:
return
new_samplerate = hparams.sampling_rate
if old_samplerate != new_samplerate:
duration = old_audio.shape[0] / old_samplerate
time_old = np.linspace(0, duration, old_audio.shape[0])
time_new = np.linspace(
0, duration,
int(old_audio.shape[0] * new_samplerate / old_samplerate))
interpolator = interpolate.interp1d(time_old, old_audio.T)
new_audio = interpolator(time_new).T.astype(np.float32)
else:
new_audio = old_audio
# Trim
wav, _ = librosa.effects.trim(new_audio,
top_db=25,
frame_length=2048,
hop_length=512)
wav = torch.from_numpy(wav).unsqueeze(0)
wav = wav.squeeze(0).numpy()
write(out_path, 22050, wav)
line = text.rstrip('\n')
return (out_path, line, speaker)
def _process_utterance(in_path, out_path, spk_name_idx):
# out_path = out_path.replace('pcm', 'wav')
out_path = out_path.replace('wav', 'wav')
dir = os.path.dirname(out_path)
# wav is saved as int 16
command = 'sox -L -c 1 -e signed -b 16 -t raw -r 44100 {} -c 1 -e signed -b 16 -t wav -r 22050 {}'\
.format(in_path, out_path)
os.system(command)
# int16 is converted into float32 here
sampling_rate, audio = read_wav_np(out_path)
wav, _ = librosa.effects.trim(audio, top_db=25, frame_length=2048, hop_length=512)
wav = torch.from_numpy(wav).unsqueeze(0)
wav = wav.squeeze(0).numpy()
# txt_file = in_path.replace('raw', 'script').replace('.pcm', '.pron')
txt_file = in_path.replace('wav', 'script').replace('.pcm', '.pron')
with open(txt_file, 'r', encoding='utf-8-sig') as f:
line = f.readline()
speaker = in_path.split('/')[spk_name_idx]
write(out_path, 22050, wav)
return (out_path, line.rstrip('\n'), speaker)
def _process_utterance(in_path):
try:
sr, wav = read_wav_np(in_path)
except:
return in_path + '\n'
return ''