def compute_input_seq(self, num_workers=0):
"""compute input sequences separately. Call it before
passing dataset to data loader."""
if not self.use_phonemes:
if self.verbose:
print(" | > Computing input sequences ...")
for idx, item in enumerate(tqdm.tqdm(self.items)):
text, *_ = item
sequence = np.asarray(text_to_sequence(
text, [self.cleaners],
tp=self.tp,
add_blank=self.add_blank),
dtype=np.int32)
self.items[idx][0] = sequence
else:
func_args = [
self.phoneme_cache_path, self.enable_eos_bos, self.cleaners,
self.phoneme_language, self.tp, self.add_blank
]
if self.verbose:
print(" | > Computing phonemes ...")
if num_workers == 0:
for idx, item in enumerate(tqdm.tqdm(self.items)):
phonemes = self._phoneme_worker([item, func_args])
self.items[idx][0] = phonemes
else:
with Pool(num_workers) as p:
phonemes = list(
tqdm.tqdm(p.imap(MyDataset._phoneme_worker,
[[item, func_args]
for item in self.items]),
total=len(self.items)))
for idx, p in enumerate(phonemes):
self.items[idx][0] = p
def load_data(self, idx):
item = self.items[idx]
if len(item) == 4:
text, wav_file, speaker_name, attn_file = item
else:
text, wav_file, speaker_name = item
attn = None
wav = np.asarray(self.load_wav(wav_file), dtype=np.float32)
if self.use_phonemes:
text = self._load_or_generate_phoneme_sequence(wav_file, text)
else:
text = np.asarray(text_to_sequence(text, [self.cleaners],
tp=self.tp,
add_blank=self.add_blank),
dtype=np.int32)
assert text.size > 0, self.items[idx][1]
assert wav.size > 0, self.items[idx][1]
if "attn_file" in locals():
attn = np.load(attn_file)
sample = {
'text': text,
'wav': wav,
'attn': attn,
'item_idx': self.items[idx][1],
'speaker_name': speaker_name,
'wav_file_name': os.path.basename(wav_file)
}
return sample
def load_data(self, idx):
item = self.items[idx]
if len(item) == 4:
text, wav_file, speaker_name, attn_file = item
else:
text, wav_file, speaker_name = item
attn = None
wav = np.asarray(self.load_wav(wav_file), dtype=np.float32)
# apply noise for augmentation
if self.use_noise_augment:
wav = wav + (1.0 / 32768.0) * np.random.rand(*wav.shape)
if not self.input_seq_computed:
if self.use_phonemes:
text = self._load_or_generate_phoneme_sequence(
wav_file,
text,
self.phoneme_cache_path,
self.enable_eos_bos,
self.cleaners,
self.phoneme_language,
self.tp,
self.add_blank,
)
else:
text = np.asarray(text_to_sequence(text, [self.cleaners],
tp=self.tp,
add_blank=self.add_blank),
dtype=np.int32)
assert text.size > 0, self.items[idx][1]
assert wav.size > 0, self.items[idx][1]
if "attn_file" in locals():
attn = np.load(attn_file)
if len(text) > self.max_seq_len:
# return a different sample if the phonemized
# text is longer than the threshold
# TODO: find a better fix
return self.load_data(100)
sample = {
"text": text,
"wav": wav,
"attn": attn,
"item_idx": self.items[idx][1],
"speaker_name": speaker_name,
"wav_file_name": os.path.basename(wav_file),
}
return sample
def compute_input_seq(self, num_workers=0):
"""Compute the input sequences with multi-processing.
Call it before passing dataset to the data loader to cache the input sequences for faster data loading."""
if not self.use_phonemes:
if self.verbose:
print(" | > Computing input sequences ...")
for idx, item in enumerate(tqdm.tqdm(self.items)):
text, *_ = item
sequence = np.asarray(
text_to_sequence(
text,
[self.cleaners],
custom_symbols=self.custom_symbols,
tp=self.characters,
add_blank=self.add_blank,
),
dtype=np.int32,
)
self.items[idx][0] = sequence
else:
func_args = [
self.phoneme_cache_path,
self.enable_eos_bos,
self.cleaners,
self.phoneme_language,
self.custom_symbols,
self.characters,
self.add_blank,
]
if self.verbose:
print(" | > Computing phonemes ...")
if num_workers == 0:
for idx, item in enumerate(tqdm.tqdm(self.items)):
phonemes = self._phoneme_worker([item, func_args])
self.items[idx][0] = phonemes
else:
with Pool(num_workers) as p:
phonemes = list(
tqdm.tqdm(
p.imap(TTSDataset._phoneme_worker,
[[item, func_args] for item in self.items]),
total=len(self.items),
))
for idx, p in enumerate(phonemes):
self.items[idx][0] = p
def load_data(self, idx):
text, wav_file, speaker_name = self.items[idx]
wav = np.asarray(self.load_wav(wav_file), dtype=np.float32)
if self.use_phonemes:
text = self._load_or_generate_phoneme_sequence(wav_file, text)
else:
text = np.asarray(text_to_sequence(text, [self.cleaners],
tp=self.tp),
dtype=np.int32)
assert text.size > 0, self.items[idx][1]
assert wav.size > 0, self.items[idx][1]
sample = {
'text': text,
'wav': wav,
'item_idx': self.items[idx][1],
'speaker_name': speaker_name
}
return sample
def load_data(self, idx):
item = self.items[idx]
if len(item) == 5:
text, wav_file, speaker_name, language_name, attn_file = item
else:
text, wav_file, speaker_name, language_name = item
attn = None
raw_text = text
wav = np.asarray(self.load_wav(wav_file), dtype=np.float32)
# apply noise for augmentation
if self.use_noise_augment:
wav = wav + (1.0 / 32768.0) * np.random.rand(*wav.shape)
if not self.input_seq_computed:
if self.use_phonemes:
text = self._load_or_generate_phoneme_sequence(
wav_file,
text,
self.phoneme_cache_path,
self.enable_eos_bos,
self.cleaners,
language_name if language_name else self.phoneme_language,
self.custom_symbols,
self.characters,
self.add_blank,
)
else:
text = np.asarray(
text_to_sequence(
text,
[self.cleaners],
custom_symbols=self.custom_symbols,
tp=self.characters,
add_blank=self.add_blank,
),
dtype=np.int32,
)
assert text.size > 0, self.items[idx][1]
assert wav.size > 0, self.items[idx][1]
if "attn_file" in locals():
attn = np.load(attn_file)
if len(text) > self.max_seq_len:
# return a different sample if the phonemized
# text is longer than the threshold
# TODO: find a better fix
return self.load_data(self.rescue_item_idx)
pitch = None
if self.compute_f0:
pitch = self.pitch_extractor.load_or_compute_pitch(
self.ap, wav_file, self.f0_cache_path)
pitch = self.pitch_extractor.normalize_pitch(
pitch.astype(np.float32))
sample = {
"raw_text": raw_text,
"text": text,
"wav": wav,
"pitch": pitch,
"attn": attn,
"item_idx": self.items[idx][1],
"speaker_name": speaker_name,
"language_name": language_name,
"wav_file_name": os.path.basename(wav_file),
}
return sample