mlobj/music_source_separation

1import argparse2import glob3 4import numpy as np5import librosa6from essentia.standard import (NSGConstantQ, 7    NSGIConstantQ)8 9import hparams10import utils11 12def parse_files(path, source):13 14  if source == 'mixture':15    path = path + 'Mixtures/Dev/*/' + str(source) + '.wav'16    paths = sorted(glob.glob(path))17  else:18    path = path + 'Sources/Dev/*/' + str(source) + '.wav'19    paths = sorted(glob.glob(path))20  return paths21 22def forward_transform(y, min_f, max_f, bpo, gamma):23  # Parameters24  params = {25            # Backward transform needs to know the signal size.26            'inputSize': y.size,27            'minFrequency': min_f,28            'maxFrequency': max_f,29            'binsPerOctave': bpo,30            # Minimum number of FFT bins per CQ channel.31            'minimumWindow': 4,32            'gamma': gamma33          }34 35 36  # Forward and backward transforms37  constantq, dcchannel, nfchannel = NSGConstantQ(**params)(y)38 39  return constantq, dcchannel, nfchannel40 41def backward_transform(c, dc, nf, orig_size, min_f, max_f, bpo, gamma):42  # Parameters43  params = {44            # Backward transform needs to know the signal size.45            'inputSize': orig_size,46            'minFrequency': min_f,47            'maxFrequency': max_f,48            'binsPerOctave': bpo,49            # Minimum number of FFT bins per CQ channel.50            'minimumWindow': 4,51            'gamma': gamma52          }53 54 55  # Forward and backward transforms56  y = NSGIConstantQ(**params)(c, dc, nf)57 58  return y59 60 61def make_chunks(c):62  cqt = np.abs(c).astype(np.float16)63  cqt = np.asfortranarray(cqt)64  padded_cqt = librosa.util.fix_length(cqt,hparams.chunk_size*np.ceil(cqt.shape[-1]/hparams.chunk_size).astype(int))65  framed_cqt = librosa.util.frame(padded_cqt,hparams.chunk_size,hparams.chunk_size)66  samples = np.transpose(framed_cqt,(2,0,1))67  cqt_input = np.expand_dims(samples,-1)68  return cqt_input69 70if __name__ == '__main__':71  args = argparse.ArgumentParser()72 73  args.add_argument('Path',metavar='path',type=str,help='Path to DSD100')74  args.add_argument('Source',metavar='source',type=str,help='Desired source to preprocess for separation. Use mixture to preprocess the mixtures')75  args.add_argument('Output_path',metavar='output_path',type=str,help='Output path for the pikled spectrograms')76 77  args = args.parse_args()78  path = args.Path79  source = args.Source80  outpath = args.Output_path81 82  if path[-1] != '/':83    path = path + '/'84  if outpath[-1] != '/':85    outpath = outpath + '/'86 87 88  files = parse_files(path, source)89  mag_lf_array = []90  mag_hf_array = []91 92  for i in range(0,len(files)):93    print(files[i])94    y, sr = librosa.load(files[i], hparams.sr, mono = True)95    C_lf,_,_ = forward_transform(y,hparams.lf_params['min_f'],hparams.lf_params['max_f'],hparams.lf_params['bins_per_octave'], hparams.lf_params['gamma'])96    C_hf,_,_ = forward_transform(y,hparams.hf_params['min_f'],hparams.hf_params['max_f'],hparams.hf_params['bins_per_octave'], hparams.hf_params['gamma'])97    c_lf = make_chunks(C_lf)98    c_hf = make_chunks(C_hf)99    mag_lf_array.append(c_lf)100    mag_hf_array.append(c_hf)101    if  i == 1:102      break103 104 105  mag_lf = utils.list_to_array(mag_lf_array)106  mag_hf = utils.list_to_array(mag_hf_array)107 108 109  filename_lf = source + '_lf.npy'110  filename_hf = source + '_hf.npy'111  utils.pickle(mag_lf, outpath, filename_lf)112  utils.pickle(mag_hf, outpath, filename_hf)113 114 115 116