techscientist / audio_preprocessor

To create a HDF file with representations for large-scale audio data

audio_preprocessor

Under development now. Open for any suggestions.

Goal

To create a single HDF file,

• that contains all the transforms of the audio files,
  • that are in a certain folder,
• where the transforms can be specified w.r.t librosa
• and so as the extensions of audio files
• with a random order for training,
• using multiprocessing

and I'm gonna use it for the training of my neural networks.

The structure of HDF

If melgram, cqt, stft have been acquired, there would be f.items() = ['melgram', 'cqt', 'stft']. Each dataset would be 4-dim numpy array, e.g. f['melgram'].shape = (1000, 1, 128, 200) when there is 1000 songs, 1 channel, n_mels=128, n_frame=200 (same is the formatting in Theano).

How to use

1. Create or modify settings.json
2. As in the example.py,

import audio_preprocessor

nogada = audio_preprocessor.Audio_Preprocessor(settings_path='settings.json')
nogada.index()
nogada.get_permutations()
nogada.convert_all()

Credits

Test music items are from http://www.bensound.com