process_data.py

import os
import pickle
from glob import iglob
import numpy as np
import librosa

DATA_AUDIO_DIR = './audio'
TARGET_SR = 8000
OUTPUT_DIR = './output'
OUTPUT_DIR_TRAIN = os.path.join(OUTPUT_DIR, 'train')
OUTPUT_DIR_TEST = os.path.join(OUTPUT_DIR, 'test')
AUDIO_LENGTH = 10000

class_ids = {
    'normal': 0,
    'murmur': 1,
    'extrahls': 2,
    'artifact': 3,
    'unlabelled': 4,
}


def extract_class_id(wav_filename):
    if 'normal' in wav_filename:
        return class_ids.get('normal')
    elif 'murmur' in wav_filename:
        return class_ids.get('murmur')
    elif 'extrahls' in wav_filename:
        return class_ids.get('extrahls')
    elif 'artifact' in wav_filename:
        return class_ids.get('artifact')
    elif 'unlabelled' in wav_filename:
        return class_ids.get('unlabelled')
    else:
        return class_ids.get('unlabelled')


def read_audio_from_filename(filename, target_sr):
    audio, _ = librosa.load(filename, sr=target_sr, mono=True)
    audio = audio.reshape(-1, 1)
    return audio


def convert_data():
    for i, wav_filename in enumerate(iglob(os.path.join(DATA_AUDIO_DIR, '**/**.wav'), recursive=True)):
        class_id = extract_class_id(wav_filename)
        audio_buf = read_audio_from_filename(wav_filename, target_sr=TARGET_SR)
        # normalize mean 0, variance 1
        audio_buf = (audio_buf - np.mean(audio_buf)) / np.std(audio_buf)
        original_length = len(audio_buf)
        print(i, wav_filename, original_length, np.round(np.mean(audio_buf), 4), np.std(audio_buf))
        if original_length < AUDIO_LENGTH:
            audio_buf = np.concatenate((audio_buf, np.zeros(shape=(AUDIO_LENGTH - original_length, 1))))
            print('PAD New length =', len(audio_buf))
        elif original_length > AUDIO_LENGTH:
            audio_buf = audio_buf[0:AUDIO_LENGTH]
            print('CUT New length =', len(audio_buf))

        output_folder = OUTPUT_DIR_TRAIN
        if i // 50 == 0:
            output_folder = OUTPUT_DIR_TEST

        output_filename = os.path.join(output_folder, str(i) + '.pkl')

        out = {'class_id': class_id,
               'audio': audio_buf,
               'sr': TARGET_SR}
        with open(output_filename, 'wb') as w:
            pickle.dump(out, w)


if __name__ == '__main__':
    convert_data()