def __init__(self,
dataset_path,
audiopaths_and_text,
text_cleaners,
n_mel_channels,
symbol_set='english_basic',
n_speakers=1,
load_mel_from_disk=True,
max_wav_value=None,
sampling_rate=None,
filter_length=None,
hop_length=None,
win_length=None,
mel_fmin=None,
mel_fmax=None,
**ignored):
self.audiopaths_and_text = load_filepaths_and_text(
dataset_path, audiopaths_and_text,
has_speakers=(n_speakers > 1))
self.load_mel_from_disk = load_mel_from_disk
if not load_mel_from_disk:
self.max_wav_value = max_wav_value
self.sampling_rate = sampling_rate
self.stft = layers.TacotronSTFT(
filter_length, hop_length, win_length,
n_mel_channels, sampling_rate, mel_fmin, mel_fmax)
def __init__(self, dataset_path, audiopaths_and_text, args):
self.audiopaths_and_text = load_filepaths_and_text(dataset_path, audiopaths_and_text)
self.max_wav_value = args.max_wav_value
self.sampling_rate = args.sampling_rate
self.stft = layers.TacotronSTFT(
args.filter_length, args.hop_length, args.win_length,
args.n_mel_channels, args.sampling_rate, args.mel_fmin,
args.mel_fmax)
self.segment_length = args.segment_length
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
def __init__(self, dataset_path, audiopaths_and_text, args, load_mel_from_disk=True):
self.audiopaths_and_text = load_filepaths_and_text(dataset_path, audiopaths_and_text)
self.text_cleaners = args.text_cleaners
self.load_mel_from_disk = load_mel_from_disk
if not load_mel_from_disk:
self.max_wav_value = args.max_wav_value
self.sampling_rate = args.sampling_rate
self.stft = layers.TacotronSTFT(
args.filter_length, args.hop_length, args.win_length,
args.n_mel_channels, args.sampling_rate, args.mel_fmin,
args.mel_fmax)
def __init__(self, audiopaths_and_text, args):
self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
self.text_cleaners = args.text_cleaners
self.max_wav_value = args.max_wav_value
self.sampling_rate = args.sampling_rate
self.load_mel_from_disk = args.load_mel_from_disk
self.stft = layers.TacotronSTFT(args.filter_length, args.hop_length,
args.win_length, args.n_mel_channels,
args.sampling_rate, args.mel_fmin,
args.mel_fmax)
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
def __init__(self, dataset_path, audiopaths_and_text, segment_length,
n_mel_channels, max_wav_value, sampling_rate, filter_length,
hop_length, win_length, mel_fmin, mel_fmax, args):
self.audiopaths_and_text = load_filepaths_and_text(
dataset_path, audiopaths_and_text)
self.max_wav_value = max_wav_value
self.sampling_rate = sampling_rate
self.stft = layers.TacotronSTFT(filter_length, hop_length, win_length,
n_mel_channels, sampling_rate,
mel_fmin, mel_fmax)
self.segment_length = segment_length
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
def __init__(self, data_utterance_paths, hparams):
"""Data loader for the PPG->Mel task.
Args:
data_utterance_paths: A text file containing a list of file paths.
hparams: The hyper-parameters.
"""
self.data_utterance_paths = load_filepaths(data_utterance_paths)
self.max_wav_value = hparams.max_wav_value
self.sampling_rate = hparams.sampling_rate
self.is_full_ppg = hparams.is_full_ppg
self.is_append_f0 = hparams.is_append_f0
self.is_cache_feats = hparams.is_cache_feats
self.load_feats_from_disk = hparams.load_feats_from_disk
self.feats_cache_path = hparams.feats_cache_path
self.ppg_subsampling_factor = hparams.ppg_subsampling_factor
self.ppg_deps = DependenciesPPG()
self.encoder_model_fpath = hparams.encoder_model_fpath
if self.is_cache_feats and self.load_feats_from_disk:
raise ValueError('If you are loading feats from the disk, do not '
'rewrite them back!')
self.stft = layers.TacotronSTFT(
hparams.filter_length, hparams.hop_length, hparams.win_length,
hparams.n_acoustic_feat_dims, hparams.sampling_rate,
hparams.mel_fmin, hparams.mel_fmax)
random.seed(hparams.seed)
random.shuffle(self.data_utterance_paths)
self.ppg_sequences = []
self.acoustic_sequences = []
self.dvec_sequences = []
if self.load_feats_from_disk:
print('Loading data from %s.' % self.feats_cache_path)
with open(self.feats_cache_path, 'rb') as f:
data = pickle.load(f)
self.ppg_sequences = data[0]
self.acoustic_sequences = data[1]
self.dvec_sequences = data[2]
else:
for utterance_path in self.data_utterance_paths:
ppg_feat_pair_spkr = self.extract_utterance_feats_spkr(utterance_path, self.is_full_ppg)
self.ppg_sequences.append(ppg_feat_pair_spkr[0].astype(np.float32))
self.acoustic_sequences.append(ppg_feat_pair_spkr[1])
self.dvec_sequences.append(ppg_feat_pair_spkr[2])
if self.is_cache_feats:
print('Caching data to %s.' % self.feats_cache_path)
with open(self.feats_cache_path, 'wb') as f:
pickle.dump([self.ppg_sequences, self.acoustic_sequences, self.dvec_sequences], f)
def __init__(self, data_utterance_paths, cache_path, hparams, bs, loop):
"""Data loader for the PPG->Mel task.
Args:
data_utterance_paths: A text file containing a list of file paths.
hparams: The hyper-parameters.
"""
self.data_utterance_paths = load_filepaths(data_utterance_paths)
self.max_wav_value = 32768.0
self.sampling_rate = hparams.sampling_rate
self.is_full_ppg = True
self.is_append_f0 = False
self.is_cache_feats = True
self.feats_cache_path = cache_path
self.ppg_subsampling_factor = 1
self.ppg_deps = DependenciesPPG()
# 20 data = n(4) * b(5)
self.n = int(bs) - 1
self.b = 5
self.l = int(loop) - 1
self.stft = layers.TacotronSTFT(
hparams.filter_length, hparams.hop_length, hparams.win_length,
hparams.n_acoustic_feat_dims, self.sampling_rate,
hparams.mel_fmin, hparams.mel_fmax)
random.seed(hparams.seed)
random.shuffle(self.data_utterance_paths)
if self.n > 0:
with open(self.feats_cache_path, 'rb') as f:
data = pickle.load(f)
self.ppg_sequences = data[0]
self.acoustic_sequences = data[1]
else:
self.ppg_sequences = []
self.acoustic_sequences = []
for utterance_path in self.data_utterance_paths[self.n * self.b + self.l * 20 : (self.n+1) * self.b + self.l * 20]:
ppg_feat_pair = self.extract_utterance_feats(utterance_path,
self.is_full_ppg)
self.ppg_sequences.append(ppg_feat_pair[0].astype(
np.float32))
self.acoustic_sequences.append(ppg_feat_pair[1])
if self.is_cache_feats:
print('Caching data to %s.' % self.feats_cache_path)
with open(self.feats_cache_path, 'wb+') as f:
pickle.dump([self.ppg_sequences, self.acoustic_sequences], f)
def __init__(self, dataset_path, audiopaths_and_text, args, speaker_ids=None):
self.audiopaths_and_text = load_filepaths_and_text(dataset_path, audiopaths_and_text)
self.text_cleaners = args.text_cleaners
self.max_wav_value = args.max_wav_value
self.sampling_rate = args.sampling_rate
self.load_mel_from_disk = args.load_mel_from_disk
self.stft = layers.TacotronSTFT(
args.filter_length, args.hop_length, args.win_length,
args.n_mel_channels, args.sampling_rate, args.mel_fmin,
args.mel_fmax)
random.seed(1234)
random.shuffle(self.audiopaths_and_text)
self.speaker_ids = speaker_ids
if speaker_ids is None:
self.speaker_ids = self.create_speaker_lookup_table(self.audiopaths_and_text)
def __init__(self, data_utterance_paths, hparams):
"""Data loader for the PPG->Mel task.
Args:
data_utterance_paths: A text file containing a list of file paths.
hparams: The hyper-parameters.
"""
self.data_utterance_paths = data_utterance_paths
self.load_feats_from_disk = hparams.load_feats_from_disk
self.max_wav_value = hparams.max_wav_value
self.sampling_rate = hparams.sampling_rate
self.is_full_ppg = hparams.is_full_ppg
self.is_append_f0 = hparams.is_append_f0
self.ppg_subsampling_factor = hparams.ppg_subsampling_factor
self.ppg_deps = DependenciesPPG()
self.stft = layers.TacotronSTFT(hparams.filter_length,
hparams.hop_length, hparams.win_length,
hparams.n_acoustic_feat_dims,
hparams.sampling_rate,
hparams.mel_fmin, hparams.mel_fmax)
random.seed(hparams.seed)
# 从disk加载数据对
self.ppg_sequences = []
self.acoustic_sequences = []
if self.load_feats_from_disk:
fpath = self.data_utterance_paths
with open(fpath, 'rb') as f:
data = pickle.load(f)
self.ppg_sequences = data[0]
self.acoustic_sequences = data[1]
print("Number of data:", len(self.ppg_sequences))
else:
print("No data!")
def __init__(
self,
dataset_path,
audiopaths_and_text,
text_cleaners,
n_mel_channels,
symbol_set='english_basic',
p_arpabet=1.0,
n_speakers=1,
load_mel_from_disk=True,
load_pitch_from_disk=True,
pitch_mean=214.72203, # LJSpeech defaults
pitch_std=65.72038,
max_wav_value=None,
sampling_rate=None,
filter_length=None,
hop_length=None,
win_length=None,
mel_fmin=None,
mel_fmax=None,
prepend_space_to_text=False,
append_space_to_text=False,
pitch_online_dir=None,
betabinomial_online_dir=None,
use_betabinomial_interpolator=True,
pitch_online_method='pyin',
**ignored):
# Expect a list of filenames
if type(audiopaths_and_text) is str:
audiopaths_and_text = [audiopaths_and_text]
self.dataset_path = dataset_path
self.audiopaths_and_text = load_filepaths_and_text(
dataset_path, audiopaths_and_text, has_speakers=(n_speakers > 1))
self.load_mel_from_disk = load_mel_from_disk
if not load_mel_from_disk:
self.max_wav_value = max_wav_value
self.sampling_rate = sampling_rate
self.stft = layers.TacotronSTFT(filter_length, hop_length,
win_length, n_mel_channels,
sampling_rate, mel_fmin, mel_fmax)
self.load_pitch_from_disk = load_pitch_from_disk
self.prepend_space_to_text = prepend_space_to_text
self.append_space_to_text = append_space_to_text
assert p_arpabet == 0.0 or p_arpabet == 1.0, (
'Only 0.0 and 1.0 p_arpabet is currently supported. '
'Variable probability breaks caching of betabinomial matrices.')
self.tp = TextProcessing(symbol_set,
text_cleaners,
p_arpabet=p_arpabet)
self.n_speakers = n_speakers
self.pitch_tmp_dir = pitch_online_dir
self.f0_method = pitch_online_method
self.betabinomial_tmp_dir = betabinomial_online_dir
self.use_betabinomial_interpolator = use_betabinomial_interpolator
if use_betabinomial_interpolator:
self.betabinomial_interpolator = BetaBinomialInterpolator()
expected_columns = (2 + int(load_pitch_from_disk) + (n_speakers > 1))
assert not (load_pitch_from_disk and self.pitch_tmp_dir is not None)
if len(self.audiopaths_and_text[0]) < expected_columns:
raise ValueError(
f'Expected {expected_columns} columns in audiopaths file. '
'The format is <mel_or_wav>|[<pitch>|]<text>[|<speaker_id>]')
if len(self.audiopaths_and_text[0]) > expected_columns:
print('WARNING: Audiopaths file has more columns than expected')
to_tensor = lambda x: torch.Tensor([x]) if type(x) is float else x
self.pitch_mean = to_tensor(pitch_mean)
self.pitch_std = to_tensor(pitch_std)
