def collect_features(self,file_path):
"""PyWorld analysis"""
sr = 16000
save_path = os.path.join(self.preprocess_dir, self.speaker, os.path.basename(file_path))
if os.path.exists(save_path):
features = np.load(save_path)
else:
wav, _ = librosa.load(file_path, sr=sr, mono=True)
wav_padded = wav_padding(wav, sr=sr, frame_period=5, multiple=4)
f0, _, sp, ap = world_decompose(wav_padded,sr)
mcep = world_encode_spectral_envelop(sp, sr, dim=24)
# Extending to 2D to stack and log zeroes 1e-16. TODO: Better solution for this
f0 = np.ma.log(f0[:,None])
#f0[f0 == -np.inf] = 1e-16
features = np.hstack((f0, mcep, ap))
features.dump(save_path)
return features
def load_wav(wavfile, sr=16000):
wav, _ = librosa.load(wavfile, sr=sr, mono=True)
return wav_padding(wav, sr=sr, frame_period=5, multiple=4)
}
print('-------------------------------')
wavs = os.listdir(data_root)
self.wavs = [os.path.join(data_root, wav_name) for wav_name in wavs]
self.samples = []
self.samples_paths = []
for wav_path in self.wavs:
_sample = wav_loader(wav_path, period=3, stride=1)
self.samples.extend(_sample)
self.samples_paths.extend([wav_path for _ in range(len(_sample))])
def __len__(self):
return len(self.samples)
def __getitem__(self, idx):
ffimg = self.samples[idx]
_path = self.samples_paths[idx]
return ffimg, _path
if __name__ == '__main__':
import cv2
import utils
wavDataset = WavDataset(is_train=True)
for _ff, _label, _path in wavDataset:
print(_ff.shape, _label, _path)
pa_wav = utils.wav_padding([_ff, _ff])
print(pa_wav.shape)
# cv2.imshow('', pa_wav[0])
# cv2.waitKey()