def __init__(self, metadata_file_path, hparams):
# self.audiopaths_and_text = load_filepaths_and_text(metadata_file_path)
self._metadata = _read_meta_yyh(metadata_file_path)
self.mel_dir = hparams.mel_dir
if hparams.use_phone:
from text.phones import Phones
self.phone_class = Phones(hparams.phone_set_file)
self.text_to_sequence = self.phone_class.text_to_sequence
else:
from text import text_to_sequence
self.text_to_sequence = text_to_sequence
self.text_cleaners = hparams.text_cleaners
self.is_multi_speakers = hparams.is_multi_speakers
self.is_multi_styles = hparams.is_multi_styles
# random.seed(1234)
random.shuffle(self._metadata)
self.utt_list = []
self.utt_data = {}
self.utt_mels = []
for i in range(len(self._metadata)):
item = self._metadata[i]
self.utt_list.append(
{"mel_frame": int(item.strip().split('|')[-1])})
self.utt_data[str(i)] = {
"text":
torch.IntTensor(
self.text_to_sequence(item.strip().split('|')[5],
self.text_cleaners)),
"mel":
torch.from_numpy(
np.load(self.mel_dir + 'out_' +
item.strip().split('|')[0] + '.npy')) * 8.0 - 4.0,
"dur":
DurationCalculator._calculate_duration(
torch.from_numpy(
np.load(self.mel_dir + 'encdec_' +
item.strip().split('|')[0] + '.npy'))),
"speaker":
int(item.strip().split('|')[4])
if self.is_multi_speakers else None,
"style":
int(item.strip().split('|')[2])
if self.is_multi_styles else None
}
# self.utt_list.append({"mel_frame": int(np.load(self.mel_dir + 'out_' + item.strip().split('|')[0] + '.npy').shape[0])})
if hparams.is_refine_style:
self.utt_mels.append((torch.from_numpy(
np.load(self.mel_dir + 'out_' +
item.strip().split('|')[0] + '.npy')) * 8.0 -
4.0).unsqueeze(0))
def get_dur(self, att_ws):
return DurationCalculator._calculate_duration(
torch.from_numpy(np.load(att_ws)))