def from_config(cls, config_manager: Config, kind: str, scan_wavs=False):
kinds = ['original', 'phonemized', 'train', 'valid']
if kind not in kinds:
raise ValueError(f'Invalid kind type. Expected one of: {kinds}')
reader = get_preprocessor_by_name(config_manager.config['data_name'])
if kind == 'train':
metadata = config_manager.train_metadata_path
elif kind == 'original':
metadata = config_manager.metadata_path
elif kind == 'valid':
metadata = config_manager.valid_metadata_path
elif kind == 'phonemized':
metadata = config_manager.phonemized_metadata_path
return cls(data_directory=config_manager.dataset_dir,
metadata_reading_function=reader,
metadata_path=metadata,
scan_wavs=scan_wavs)
def phoneme_lengths(phonemes, phoneme_padding=0):
return tf.reduce_sum(tf.cast(phonemes != phoneme_padding, tf.int32),
axis=-1)
if __name__ == '__main__':
from preprocessing.metadata_readers import get_preprocessor_by_name
from preprocessing.datasets import DataReader, AutoregressivePreprocessor, TextMelDataset
from preprocessing.text.tokenizer import Tokenizer
from preprocessing.text.symbols import all_phonemes
from pathlib import Path
ljspeech_folder = '/Volumes/data/datasets/LJSpeech-1.1'
metadata_path = '/Volumes/data/datasets/LJSpeech-1.1/phonemized_metadata.txt'
metadata_reader = get_preprocessor_by_name('ljspeech')
data_reader = DataReader(data_directory=ljspeech_folder,
metadata_path=metadata_path,
metadata_reading_function=metadata_reader,
scan_wavs=True)
mel_dir = Path('/Volumes/data/datasets/LJSpeech-1.1/mels')
tokenizer = Tokenizer(alphabet=all_phonemes)
preprocessor = AutoregressivePreprocessor(mel_channels=80,
mel_start_value=.5,
mel_end_value=-.5,
tokenizer=tokenizer)
dataset_creator = TextMelDataset(data_reader=data_reader,
preprocessor=preprocessor,
mel_directory=mel_dir)
dataset = dataset_creator.get_dataset(