utils/denoise_pipeline.py

import librosa
import torch
from demucs.apply import apply_model
from demucs.pretrained import get_model
from scipy.io.wavfile import write

demucs_model = get_model('cfa93e08')


def denoise(filename: str, device: str, out_filename='denoise.wav') -> str:
    wav_ref, sr = librosa.load(filename, mono=False, sr=44100)
    wav = torch.tensor(wav_ref)
    wav = torch.cat([wav.unsqueeze(0), wav.unsqueeze(0)]) if len(wav.shape) == 1 else wav
    ref = wav.mean(0)
    wav = (wav - ref.mean()) / wav.std()
    sources = apply_model(
        demucs_model, wav[None], device=device, shifts=1, split=True, overlap=0.1, progress=True, num_workers=0
    )[0]
    sources = sources * ref.std() + ref.mean()
    vocal_wav = sources[-1]
    vocal_wav = vocal_wav / max(1.01 * vocal_wav.abs().max(), 1)
    vocal_wav = vocal_wav.numpy()
    vocal_wav = librosa.to_mono(vocal_wav)
    vocal_wav = vocal_wav.T
    vocal_wav = librosa.resample(vocal_wav, orig_sr=44100, target_sr=48000)
    write(out_filename, 48000, vocal_wav)

    return out_filename


if __name__ == '__main__':
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    denoise(filename='../oxx.wav', device=device)