def run(audio_file: str, out_path: str, model_name: str, pretrained_path: str, lowpass_freq: int = 0):
wav, sr = librosa.load(audio_file, sr=settings.SAMPLE_RATE)
wav = preemphasis(wav)
if wav.dtype != np.float32:
wav = wav.astype(np.float32)
# load model
model = __load_model(model_name, pretrained_path)
# make tensor wav
wav = torch.FloatTensor(wav).unsqueeze(0).cuda()
# inference
print('Inference ...')
with torch.no_grad():
out_wav = model(wav)
out_wav = out_wav[0].cpu().numpy()
if lowpass_freq:
out_wav = lowpass(out_wav, frequency=lowpass_freq)
# save wav
librosa.output.write_wav(out_path, inv_preemphasis(out_wav).clip(-1., 1.), settings.SAMPLE_RATE)
print('Finish !')
def run(*audio_files: [str],
out_path: str,
model_name: str,
pretrained_path: str,
lowpass_freq: int = 0,
sample_rate: int = 22050):
for audio_file in audio_files:
print('Loading audio file...')
wav, sr = librosa.load(audio_file, sr=sample_rate)
wav = preemphasis(wav)
if wav.dtype != np.float32:
wav = wav.astype(np.float32)
# load model
model = __load_model(model_name, pretrained_path)
# make tensor wav
wav = torch.FloatTensor(wav).unsqueeze(0)
# inference
print('Inference ...')
with torch.no_grad():
out_wav = model(wav)
out_wav = out_wav[0].cpu().numpy()
if lowpass_freq:
out_wav = lowpass(out_wav, frequency=lowpass_freq)
# save wav
librosa.output.write_wav(
os.path.join(
out_path,
os.path.basename(audio_file).rsplit(".", 1)[0] + "_out.wav"),
inv_preemphasis(out_wav).clip(-1., 1.), sample_rate)
print('Finish !')