preprocess_audio.py

from scipy.io.wavfile import write
import librosa
import numpy as np
import argparse
import os

sr = 22050
max_wav_value=32768.0
trim_fft_size = 1024
trim_hop_size = 256
trim_top_db = 40

def preprocess_audio(file_list, dataset_prefix, silence_audio_size):
    for F in file_list:
        f = open(F)
        R = f.readlines()
        f.close()
        print('='*5+F+'='*5)

        for i, r in enumerate(R):
            wav_file = os.path.join(dataset_prefix, r.split('|')[0])
            data, sampling_rate = librosa.core.load(wav_file, sr)
            data = data / np.abs(data).max() *0.999
            data_= librosa.effects.trim(data, top_db= trim_top_db, frame_length=trim_fft_size, hop_length=trim_hop_size)[0]
            data_ = data_*max_wav_value
            data_ = np.append(data_, [0.]*silence_audio_size)
            data_ = data_.astype(dtype=np.int16)
            write(wav_file, sr, data_)
            #print(len(data),len(data_))
            if(i%100 == 0):
                print (i)

if __name__ == "__main__":
    """
    usage
    python preprocess_audio.py -f=namgung-highquality/metadata.csv -p=namgung-highquality/wavs -s=3
    python preprocess_audio.py -f=namgung-old/metadata.csv -p=namgung-old/wavs -s=0
    """
    parser = argparse.ArgumentParser()
    parser.add_argument('-f', '--file_list', type=str,
                        help='file list to preprocess')
    parser.add_argument('-s', '--silence_mel_padding', type=int, default=0,
                        help='silence audio size is hop_length * silence mel padding')
    parser.add_argument('-p', '--dataset_prefix', type=str,
                        help='audio path = dataset_prefix + file_prefix')
    args = parser.parse_args()
    file_list = args.file_list.split(',')
    dataset_prefix = args.dataset_prefix
    silence_audio_size = trim_hop_size * args.silence_mel_padding

    preprocess_audio(file_list, dataset_prefix, silence_audio_size)