def __getitem__(self, index):
id = self._metadata[index][4].split(".")[0]
x_ = self._metadata[index][3].split()
if self.use_phonemes:
x = phonemes_to_sequence(x_)
else:
x = text_to_sequence(x_, self.tts_cleaner_names, self.eos)
mel = np.load(f"{self.path}mels/{id}.npy")
durations = str_to_int_list(self._metadata[index][2])
e = remove_outlier(
np.load(f"{self.path}energy/{id}.npy")
) # self._norm_mean_std(np.load(f'{self.path}energy/{id}.npy'), self.e_mean, self.e_std, True)
p = remove_outlier(
np.load(f"{self.path}pitch/{id}.npy")
) # self._norm_mean_std(np.load(f'{self.path}pitch/{id}.npy'), self.f0_mean, self.f0_std, True)
mel_len = mel.shape[1]
durations = durations[:len(x)]
durations[-1] = durations[-1] + (mel.shape[1] - sum(durations))
assert mel.shape[1] == sum(durations)
return (
np.array(x),
mel.T,
id,
mel_len,
np.array(durations),
e,
p,
) # Mel [T, num_mel]
def _norm_mean_std(self, x, mean, std, is_remove_outlier = False):
if is_remove_outlier:
x = remove_outlier(x)
zero_idxs = np.where(x == 0.0)[0]
x = (x - mean) / std
x[zero_idxs] = 0.0
return x
nz_min_p = []
nz_min_e = []
energy_path = os.path.join(hp.data.data_dir, "energy")
pitch_path = os.path.join(hp.data.data_dir, "pitch")
mel_path = os.path.join(hp.data.data_dir, "mels")
energy_files = get_files(energy_path, extension=".npy")
pitch_files = get_files(pitch_path, extension=".npy")
mel_files = get_files(mel_path, extension=".npy")
assert len(energy_files) == len(pitch_files) == len(mel_files)
energy_vecs = []
for f in tqdm(energy_files):
e = np.load(f)
e = remove_outlier(e)
energy_vecs.append(e)
min_e.append(e.min())
nz_min_e.append(e[e > 0].min())
max_e.append(e.max())
nonzeros = np.concatenate([v[np.where(v != 0.0)[0]] for v in energy_vecs])
e_mean, e_std = np.mean(nonzeros), np.std(nonzeros)
print("Non zero Min Energy : {}".format(min(nz_min_e)))
print("Max Energy : {}".format(max(max_e)))
print("Energy mean : {}".format(e_mean))
print("Energy std: {}".format(e_std))
pitch_vecs = []
bad_pitch = []
for f in tqdm(pitch_files):