def run_audio(data, audio_base, audio_path, audio_conf):
if os.path.exists(audio_path):
print('%s already exists. skip' % audio_path)
return
print('Dumping audio to HDF5 : %s' % audio_path)
print(' audio_conf : %s' % audio_conf)
audio_conf['num_mel_bins'] = audio_conf.get('num_mel_bins', 40)
audio_conf['target_length'] = audio_conf.get('target_length', 2048)
audio_conf['use_raw_length'] = audio_conf.get('use_raw_length', False)
assert(not audio_conf['use_raw_length'])
# dump audio
n = len(data)
Path(os.path.dirname(audio_path)).mkdir(parents=True, exist_ok=True)
f = h5py.File(audio_path, 'w')
dset_mel_shape = (n, audio_conf['num_mel_bins'],
audio_conf['target_length'])
dset_mel = f.create_dataset('melspec', dset_mel_shape, dtype='f')
dset_len = f.create_dataset('melspec_len', (n,), dtype='i8')
start = time.time()
for i, d in enumerate(data):
y, sr = librosa.load('%s/%s' % (audio_base, d['wav']), None)
logspec, n_frames = compute_spectrogram(y, sr, audio_conf)
dset_mel[i, :, :] = logspec
dset_len[i] = n_frames
if i % 100 == 0:
t = time.time() - start
print('processed %d / %d audios (%.fs)' % (i, n, t))
def load_raw_spectrogram(path):
audio_conf = {
'audio_type': 'spectrogram',
'sample_rate': 16000,
'window_size': 0.005,
'window_stride': 0.002,
'n_fft': 1024,
'use_raw_length': True
}
y, sr = librosa.load(path, audio_conf['sample_rate'])
logspec, _ = compute_spectrogram(y, sr, audio_conf)
logspec = logspec.numpy()
return logspec, audio_conf['window_stride'], audio_conf['sample_rate']
def load_mel_spectrogram_and_path(path):
y, sr = librosa.load(path, 16000)
logmelspec, n_frames = compute_spectrogram(y, sr, audio_conf)
return logmelspec, n_frames, path
def _LoadAudio(self, path):
y, sr = librosa.load(path, None)
logspec, n_frames = compute_spectrogram(y, sr, self.audio_conf)
return logspec, n_frames