cli wrapper for https://github.com/haoheliu/voicefixer_main

voicefixer_cli.py

"""
VoiceFixer

$ python voicefixer_cli.py -i ~/Downloads/download.mp3

# for running on multiple files
$ find path/to/folder -name "*.wav"  -not -name "*denoised*" |xargs -P 20 -I{} sh -c 'python voicefixer_cli.py -i {}'

"""
"""
# installation:

sudo apt-get install libsox-fmt-all libsox-dev sox libsndfile-dev

pip install torchaudio
pip install ffmpeg-python
pip install --upgrade ddsp
pip install voicefixer==0.0.9
pip install git+https://github.com/facebookresearch/WavAugment.git
pip install note_seq
pip install tensorflow
"""

import argparse
from pathlib import Path


rate = 16000

parser = argparse.ArgumentParser(__doc__, formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-i', '--input', type=Path, help='input file path')
parser.add_argument('-o', '--out', default='{input}{mode}{noisified}_denoised.wav', help='denoised output file path')
parser.add_argument('-n', '--noisify', action='store_true', help='add noise before denoising')
parser.add_argument('-m', '--modes', nargs='+', type=int, choices={0, 1, 2}, default=[0, 1, 2])
args = parser.parse_args()

args.input = args.input.as_posix()

# Alias these for backwards compatibility and ease.
print('importing')
# import IPython.display as ipd
import augment
import librosa
import matplotlib.pyplot as plt
import numpy as np
import scipy.io
import soundfile as sf
import torch
from matplotlib import cm
from scipy.io import wavfile
from voicefixer import VoiceFixer


#@title Run this block to define some helper functions
def uniform_sample(lower, upper):
    if (abs(lower - upper) < 1e-5):
        return upper
    return float((upper - lower) * torch.rand(1) + lower)


def show_spectrogram(file_path):
    samples, _ = librosa.load(file_path, sr=44100)
    plt.figure(figsize=(12, 4))
    plt.subplot(121)
    plt.title("Linear Spectrogram")
    librosa.display.specshow(
        np.log10(np.abs(librosa.stft(samples)) + 1e-8),
        sr=44100,
        x_axis='frames',
        y_axis='linear',
        cmap=cm.jet,
        vmax=2.8,
        vmin=-1.7)
    plt.subplot(122)
    plt.title("Mel Spectrogram")
    librosa.display.specshow(
        np.log10(np.abs(librosa.feature.melspectrogram(samples, sr=44100)) + 1e-8),
        sr=44100,
        x_axis='frames',
        y_axis='mel',
        cmap=cm.jet,
        vmax=2.5,
        vmin=-4)
    plt.show()


print('loading voice filter', end=' ... ')
vf = VoiceFixer()  # Initilize a voicefixer.
print('loaded voicefixer')

sr, a = scipy.io.wavfile.read(args.input)
if len(a.shape) == 2:
    a = a[:, 0]
a = librosa.util.normalize(a.astype('float32'))
a = librosa.resample(a, sr, 16000)
a.min(), a.max()

if args.noisify:
    # Generate Random Distortion Parameters
    clipping_ratio = uniform_sample(lower=0.25, upper=1.0)
    print("clipping ratio: ", clipping_ratio)  # lower>=0.1, upper<=1.0
    lowpass_frequency = uniform_sample(lower=4000, upper=8000)
    print("lowpass cutoff frequency: ", lowpass_frequency)  # lower>=1000, and upper have no limit
    reverb_level = uniform_sample(lower=10, upper=80)  # lower>=0, upper<=100
    dumping_factor = uniform_sample(lower=10, upper=50)  # lower>=0, upper<=100
    room_size = uniform_sample(lower=10, upper=50)
    print("reverberate level, dumping_factor, room_size = ", reverb_level, dumping_factor, room_size)  # lower>=0, upper<=100
    highpass = uniform_sample(lower=10, upper=1500)
    print("highpass cutoff frequency: ", highpass)  # lower>=0, upper<=2000

    # Apply Distortion Effects
    effect_chain = augment.EffectChain().clip(clipping_ratio)  # clipping ratio
    effect_chain = effect_chain.lowpass(lowpass_frequency)  # remove high frequency information
    effect_chain = effect_chain.reverb(reverb_level, dumping_factor, room_size).channels(1)  # reverberate level, dumping factor, room size
    effect_chain = effect_chain.highpass(highpass)  # remove low frequency information
    noise_generator = lambda: torch.zeros_like(torch.tensor(a)).uniform_()
    effect_chain = effect_chain.additive_noise(noise_generator, snr=15)
    # print(effect_chain)
    y = effect_chain.apply(torch.tensor(a), src_info={'rate': rate}, target_info={'rate': rate})

    out = args.out.format(input=args.input, noisified="_noisified", mode=f'_distorted')

    print("Distorted speech file name:", out)
    sf.write(out, y[0, ...].numpy(), rate)
    # sf.write(str(second)+"_"+str(rate)+".wav",a,rate)
    # sf.write(str(second)+"_"+str(rate)+".wav",a,rate)
    # show_spectrogram(args.out.format(input=args.input, mode='distorted'))
    # ipd.Audio(args.out.format(input=args.input, mode='distorted'), rate=44100)
else:
    y = torch.Tensor(a).unsqueeze(0)

for mode in args.modes:
    out = args.out.format(input=args.input, noisified="_noisified" if args.noisify else "", mode=f'_mode{mode}')
    print("Restore: ", out)
    vf.restore(input=args.input, output=out, cuda=True, mode=mode)
    # show_spectrogram(out)
    # ipd.Audio(out, rate=44100)