def phoneme_to_sequence(text,
cleaner_names,
language,
enable_eos_bos=False,
tp=None):
# pylint: disable=global-statement
global _phonemes_to_id
if tp:
_, _phonemes = make_symbols(**tp)
_phonemes_to_id = {s: i for i, s in enumerate(_phonemes)}
sequence = []
text = text.replace(":", "")
clean_text = _clean_text(text, cleaner_names)
to_phonemes = text2phone(clean_text, language)
if to_phonemes is None:
print("!! After phoneme conversion the result is None. -- {} ".format(
clean_text))
# iterate by skipping empty strings - NOTE: might be useful to keep it to have a better intonation.
for phoneme in filter(None, to_phonemes.split('|')):
sequence += _phoneme_to_sequence(phoneme)
# Append EOS char
if enable_eos_bos:
sequence = pad_with_eos_bos(sequence, tp=tp)
return sequence
def text_to_sequence(text, cleaner_names, tp=None):
'''Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
The text can optionally have ARPAbet sequences enclosed in curly braces embedded
in it. For example, "Turn left on {HH AW1 S S T AH0 N} Street."
Args:
text: string to convert to a sequence
cleaner_names: names of the cleaner functions to run the text through
Returns:
List of integers corresponding to the symbols in the text
'''
# pylint: disable=global-statement
global _symbol_to_id
if tp:
_symbols, _ = make_symbols(**tp)
_symbol_to_id = {s: i for i, s in enumerate(_symbols)}
sequence = []
# Check for curly braces and treat their contents as ARPAbet:
while text:
m = _CURLY_RE.match(text)
if not m:
sequence += _symbols_to_sequence(_clean_text(text, cleaner_names))
break
sequence += _symbols_to_sequence(_clean_text(m.group(1),
cleaner_names))
sequence += _arpabet_to_sequence(m.group(2))
text = m.group(3)
return sequence
def pad_with_eos_bos(phoneme_sequence, tp=None):
global _PHONEMES_TO_ID, _bos, _eos
if tp:
_bos = tp['bos']
_eos = tp['eos']
_, phonemes = make_symbols(**tp)
_PHONEMES_TO_ID = {s: i for i, s in enumerate(phonemes)}
return [_PHONEMES_TO_ID[_bos]] + list(phoneme_sequence) + [_PHONEMES_TO_ID[_eos]]
def _create_random_model(self):
config = load_config(
os.path.join(get_tests_output_path(), 'dummy_model_config.json'))
if 'text' in config.keys():
symbols, phonemes = make_symbols(**config.text)
num_chars = len(phonemes) if config.use_phonemes else len(symbols)
model = setup_model(num_chars, 0, config)
output_path = os.path.join(get_tests_output_path())
save_checkpoint(model, None, None, None, output_path, 10, 10)
def pad_with_eos_bos(phoneme_sequence, tp=None):
# pylint: disable=global-statement
global _phonemes_to_id, _bos, _eos
if tp:
_bos = tp['bos']
_eos = tp['eos']
_, _phonemes = make_symbols(**tp)
_phonemes_to_id = {s: i for i, s in enumerate(_phonemes)}
return [_phonemes_to_id[_bos]
] + list(phoneme_sequence) + [_phonemes_to_id[_eos]]
def _create_random_model(self):
# pylint: disable=global-statement
global symbols, phonemes
config = load_config(
os.path.join(get_tests_output_path(), 'dummy_model_config.json'))
if 'characters' in config.keys():
symbols, phonemes = make_symbols(**config.characters)
num_chars = len(phonemes) if config.use_phonemes else len(symbols)
model = setup_model(num_chars, 0, config)
output_path = os.path.join(get_tests_output_path())
save_checkpoint(model, None, None, None, output_path, 10, 10)
def sequence_to_phoneme(sequence, tp=None):
'''Converts a sequence of IDs back to a string'''
global _ID_TO_PHONEMES
if tp:
_, phonemes = make_symbols(**tp)
_ID_TO_PHONEMES = {i: s for i, s in enumerate(phonemes)}
for symbol_id in sequence:
if symbol_id in _ID_TO_PHONEMES:
s = _ID_TO_PHONEMES[symbol_id]
result += s
return result.replace('}{', ' ')
def sequence_to_phoneme(sequence, tp=None):
# pylint: disable=global-statement
'''Converts a sequence of IDs back to a string'''
global _id_to_phonemes
result = ''
if tp:
_, _phonemes = make_symbols(**tp)
_id_to_phonemes = {i: s for i, s in enumerate(_phonemes)}
for symbol_id in sequence:
if symbol_id in _id_to_phonemes:
s = _id_to_phonemes[symbol_id]
result += s
return result.replace('}{', ' ')
def sequence_to_text(sequence, tp=None):
'''Converts a sequence of IDs back to a string'''
global _ID_TO_SYMBOL
if tp:
symbols, _ = make_symbols(**tp)
_ID_TO_SYMBOL = {i: s for i, s in enumerate(symbols)}
result = ''
for symbol_id in sequence:
if symbol_id in _ID_TO_SYMBOL:
s = _ID_TO_SYMBOL[symbol_id]
# Enclose ARPAbet back in curly braces:
if len(s) > 1 and s[0] == '@':
s = '{%s}' % s[1:]
result += s
return result.replace('}{', ' ')
def sequence_to_text(sequence, tp=None):
'''Converts a sequence of IDs back to a string'''
# pylint: disable=global-statement
global _id_to_symbol
if tp:
_symbols, _ = make_symbols(**tp)
_id_to_symbol = {i: s for i, s in enumerate(_symbols)}
result = ''
for symbol_id in sequence:
if symbol_id in _id_to_symbol:
s = _id_to_symbol[symbol_id]
# Enclose ARPAbet back in curly braces:
if len(s) > 1 and s[0] == '@':
s = '{%s}' % s[1:]
result += s
return result.replace('}{', ' ')
args = parser.parse_args()
if args.vocoder_path != "":
assert args.use_cuda, " [!] Enable cuda for vocoder."
from WaveRNN.models.wavernn import Model as VocoderModel
# load the config
C = load_config(args.config_path)
C.forward_attn_mask = True
# load the audio processor
ap = AudioProcessor(**C.audio)
# if the vocabulary was passed, replace the default
if 'characters' in C.keys():
symbols, phonemes = make_symbols(**C.characters)
# load speakers
if args.speakers_json != '':
speakers = json.load(open(args.speakers_json, 'r'))
num_speakers = len(speakers)
else:
num_speakers = 0
# load the model
num_chars = len(phonemes) if C.use_phonemes else len(symbols)
model = setup_model(num_chars, num_speakers, C)
cp = torch.load(args.model_path)
model.load_state_dict(cp['model'])
model.eval()
if args.use_cuda:
def main(args): # pylint: disable=redefined-outer-name
global meta_data_train, meta_data_eval, symbols, phonemes
# Audio processor
ap = AudioProcessor(**c.audio)
print(" > TTS symbols {}".format(len(symbols)))
print(symbols)
print(" > TTS phonemes {}".format(len(phonemes)))
print(phonemes)
print('-' * 50)
# if the vocabulary was passed, replace the default
if 'text' in c.keys():
symbols, phonemes = make_symbols(**c.text)
print(" > TTS symbols {}".format(len(symbols)))
print(symbols)
print(" > TTS phonemes {}".format(len(phonemes)))
print(phonemes)
# DISTRUBUTED
if num_gpus > 1:
init_distributed(args.rank, num_gpus, args.group_id,
c.distributed["backend"], c.distributed["url"])
num_chars = len(phonemes) if c.use_phonemes else len(symbols)
# load data instances
meta_data_train, meta_data_eval = load_meta_data(c.datasets)
# parse speakers
if c.use_speaker_embedding:
speakers = get_speakers(meta_data_train)
if args.restore_path:
prev_out_path = os.path.dirname(args.restore_path)
speaker_mapping = load_speaker_mapping(prev_out_path)
assert all([speaker in speaker_mapping
for speaker in speakers]), "As of now you, you cannot " \
"introduce new speakers to " \
"a previously trained model."
else:
speaker_mapping = {name: i for i, name in enumerate(speakers)}
save_speaker_mapping(OUT_PATH, speaker_mapping)
num_speakers = len(speaker_mapping)
print("Training with {} speakers: {}".format(num_speakers,
", ".join(speakers)))
else:
num_speakers = 0
print(" | > Num chars : {}".format(num_chars))
model = setup_model(num_chars, num_speakers, c)
print(" | > Num output units : {}".format(ap.num_freq), flush=True)
params = set_weight_decay(model, c.wd)
optimizer = RAdam(params, lr=c.lr, weight_decay=0)
if c.stopnet and c.separate_stopnet:
optimizer_st = RAdam(model.decoder.stopnet.parameters(),
lr=c.lr,
weight_decay=0)
else:
optimizer_st = None
if c.loss_masking:
criterion = L1LossMasked() if c.model in ["Tacotron", "TacotronGST"
] else MSELossMasked()
else:
criterion = nn.L1Loss() if c.model in ["Tacotron", "TacotronGST"
] else nn.MSELoss()
criterion_st = nn.BCEWithLogitsLoss(
pos_weight=torch.tensor(10)) if c.stopnet else None
if args.restore_path:
checkpoint = torch.load(args.restore_path, map_location='cpu')
try:
# TODO: fix optimizer init, model.cuda() needs to be called before
# optimizer restore
# optimizer.load_state_dict(checkpoint['optimizer'])
if c.reinit_layers:
raise RuntimeError
model.load_state_dict(checkpoint['model'])
except:
print(" > Partial model initialization.")
model_dict = model.state_dict()
model_dict = set_init_dict(model_dict, checkpoint, c)
model.load_state_dict(model_dict)
del model_dict
for group in optimizer.param_groups:
group['lr'] = c.lr
print(" > Model restored from step %d" % checkpoint['step'],
flush=True)
args.restore_step = checkpoint['step']
else:
args.restore_step = 0
if use_cuda:
model.cuda()
criterion.cuda()
if criterion_st:
criterion_st.cuda()
# DISTRUBUTED
if num_gpus > 1:
model = apply_gradient_allreduce(model)
if c.noam_schedule:
scheduler = NoamLR(optimizer,
warmup_steps=c.warmup_steps,
last_epoch=args.restore_step - 1)
else:
scheduler = None
num_params = count_parameters(model)
print("\n > Model has {} parameters".format(num_params), flush=True)
if 'best_loss' not in locals():
best_loss = float('inf')
global_step = args.restore_step
for epoch in range(0, c.epochs):
# set gradual training
if c.gradual_training is not None:
r, c.batch_size = gradual_training_scheduler(global_step, c)
c.r = r
model.decoder.set_r(r)
if c.bidirectional_decoder:
model.decoder_backward.set_r(r)
print(" > Number of outputs per iteration:", model.decoder.r)
train_loss, global_step = train(model, criterion, criterion_st,
optimizer, optimizer_st, scheduler, ap,
global_step, epoch)
val_loss = evaluate(model, criterion, criterion_st, ap, global_step,
epoch)
print(" | > Training Loss: {:.5f} Validation Loss: {:.5f}".format(
train_loss, val_loss),
flush=True)
target_loss = train_loss
if c.run_eval:
target_loss = val_loss
best_loss = save_best_model(model, optimizer, target_loss, best_loss,
OUT_PATH, global_step, epoch)
def main(args): # pylint: disable=redefined-outer-name
# pylint: disable=global-variable-undefined
global meta_data_train, meta_data_eval, symbols, phonemes
# Audio processor
ap = AudioProcessor(**c.audio)
if 'characters' in c.keys():
symbols, phonemes = make_symbols(**c.characters)
# DISTRUBUTED
if num_gpus > 1:
init_distributed(args.rank, num_gpus, args.group_id,
c.distributed["backend"], c.distributed["url"])
num_chars = len(phonemes) if c.use_phonemes else len(symbols)
# load data instances
meta_data_train, meta_data_eval = load_meta_data(c.datasets)
# parse speakers
if c.use_speaker_embedding:
speakers = get_speakers(meta_data_train)
if args.restore_path:
prev_out_path = os.path.dirname(args.restore_path)
speaker_mapping = load_speaker_mapping(prev_out_path)
assert all([speaker in speaker_mapping
for speaker in speakers]), "As of now you, you cannot " \
"introduce new speakers to " \
"a previously trained model."
else:
speaker_mapping = {name: i for i, name in enumerate(speakers)}
save_speaker_mapping(OUT_PATH, speaker_mapping)
num_speakers = len(speaker_mapping)
print("Training with {} speakers: {}".format(num_speakers,
", ".join(speakers)))
else:
num_speakers = 0
model = setup_model(num_chars, num_speakers, c)
print(" | > Num output units : {}".format(ap.num_freq), flush=True)
params = set_weight_decay(model, c.wd)
optimizer = RAdam(params, lr=c.lr, weight_decay=0)
if c.stopnet and c.separate_stopnet:
optimizer_st = RAdam(model.decoder.stopnet.parameters(),
lr=c.lr,
weight_decay=0)
else:
optimizer_st = None
# setup criterion
criterion = TacotronLoss(c, stopnet_pos_weight=10.0, ga_sigma=0.4)
if args.restore_path:
checkpoint = torch.load(args.restore_path, map_location='cpu')
try:
# TODO: fix optimizer init, model.cuda() needs to be called before
# optimizer restore
# optimizer.load_state_dict(checkpoint['optimizer'])
if c.reinit_layers:
raise RuntimeError
model.load_state_dict(checkpoint['model'])
except:
print(" > Partial model initialization.")
model_dict = model.state_dict()
model_dict = set_init_dict(model_dict, checkpoint, c)
model.load_state_dict(model_dict)
del model_dict
for group in optimizer.param_groups:
group['lr'] = c.lr
print(" > Model restored from step %d" % checkpoint['step'],
flush=True)
args.restore_step = checkpoint['step']
else:
args.restore_step = 0
if use_cuda:
model.cuda()
criterion.cuda()
# DISTRUBUTED
if num_gpus > 1:
model = apply_gradient_allreduce(model)
if c.noam_schedule:
scheduler = NoamLR(optimizer,
warmup_steps=c.warmup_steps,
last_epoch=args.restore_step - 1)
else:
scheduler = None
num_params = count_parameters(model)
print("\n > Model has {} parameters".format(num_params), flush=True)
if 'best_loss' not in locals():
best_loss = float('inf')
global_step = args.restore_step
for epoch in range(0, c.epochs):
c_logger.print_epoch_start(epoch, c.epochs)
# set gradual training
if c.gradual_training is not None:
r, c.batch_size = gradual_training_scheduler(global_step, c)
c.r = r
model.decoder.set_r(r)
if c.bidirectional_decoder:
model.decoder_backward.set_r(r)
print("\n > Number of output frames:", model.decoder.r)
train_avg_loss_dict, global_step = train(model, criterion, optimizer,
optimizer_st, scheduler, ap,
global_step, epoch)
eval_avg_loss_dict = evaluate(model, criterion, ap, global_step, epoch)
c_logger.print_epoch_end(epoch, eval_avg_loss_dict)
target_loss = train_avg_loss_dict['avg_postnet_loss']
if c.run_eval:
target_loss = eval_avg_loss_dict['avg_postnet_loss']
best_loss = save_best_model(target_loss, best_loss, model, optimizer,
global_step, epoch, c.r, OUT_PATH)
def main(**kwargs):
global symbols, phonemes # pylint: disable=global-statement
current_date = date.today()
current_date = current_date.strftime("%B %d %Y")
start_time = time.time()
# read passed variables from gui
text = kwargs['text'] # text to generate speech from
use_cuda = kwargs['use_cuda'] # if gpu exists default is true
project = kwargs['project'] # path to project folder
vocoder_type = kwargs['vocoder'] # vocoder type, default is GL
use_gst = kwargs['use_gst'] # use style_wave for prosody
style_dict = kwargs['style_input'] # use style_wave for prosody
speaker_id = kwargs['speaker_id'] # name of the selected speaker
sentence_file = kwargs['sentence_file'] # path to file if generate from file
out_path = kwargs['out_path'] # path to save the output wav
batched_vocoder = True
# load speakers
speakers_file_path = Path(project, "speakers.json")
if speakers_file_path.is_file():
speaker_data = json.load(open(speakers_file_path, 'r'))
num_speakers = len(speaker_data)
#get the speaker id for selected speaker
if speaker_id >= num_speakers:
print('Speaker ID outside of number of speakers range. Using default 0.')
speaker_id = 0
speaker_name = [speaker for speaker, id in speaker_data.items() if speaker_id == id][0]
else:
speaker_name = [speaker for speaker, id in speaker_data.items() if speaker_id == id][0]
else:
speaker_name = 'Default'
num_speakers = 0
speaker_id = None
# load the config
config_path = Path(project, "config.json")
C = load_config(config_path)
if use_gst:
if style_dict is not None:
style_input = style_dict
else:
style_input = None
# load the audio processor
ap = AudioProcessor(**C.audio)
# if the vocabulary was passed, replace the default
if 'characters' in C.keys():
symbols, phonemes = make_symbols(**C.characters)
# find the tts model file in project folder
try:
tts_model_file = glob(str(Path(project, '*.pth.tar')))
if not tts_model_file:
raise FileNotFoundError
model_path = tts_model_file[0]
except FileNotFoundError:
print('[!] TTS Model not found in path: "{}"'.format(project))
# load the model
num_chars = len(phonemes) if C.use_phonemes else len(symbols)
model = setup_model(num_chars, num_speakers, C)
# if gpu is not available use cpu
model, state = load_checkpoint(model, model_path, use_cuda=use_cuda)
model.decoder.max_decoder_steps = 2000
model.eval()
print(' > Model step:', state['step'])
print(' > Model r: ', state['r'])
# load vocoder
if vocoder_type is 'MelGAN':
try:
model_file = glob(str(Path(project, 'vocoder/*.pth.tar')))
vocoder, ap_vocoder = load_vocoder(str(Path('TTS')),
str(model_file[0]),
str(Path(project, 'vocoder/config.json')),
use_cuda)
except Exception:
print('[!] Error loading vocoder: "{}"'.format(project))
sys.exit(0)
elif vocoder_type is 'WaveRNN':
try:
model_file = glob(str(Path(project, 'vocoder/*.pkl')))
vocoder, ap_vocoder = load_vocoder(str(Path('TTS')), str(model_file[0]), str(Path(project, 'config.yml')), use_cuda)
except Exception:
print('[!] Error loading vocoder: "{}"'.format(project))
sys.exit(0)
else:
vocoder, ap_vocoder = None, None
print(" > Vocoder: {}".format(vocoder_type))
print(' > Using style input: {}\n'.format(style_input))
if sentence_file != '':
with open(sentence_file, "r", encoding='utf8') as f:
list_of_sentences = [s.strip() for s in f.readlines()]
else:
list_of_sentences = [text.strip()]
# iterate over every passed sentence and synthesize
for _, tts_sentence in enumerate(list_of_sentences):
wav_list = []
# remove character which are not alphanumerical or contain ',. '
tts_sentence = clean_sentence(tts_sentence)
print(" > Text: {}".format(tts_sentence))
# build filename
current_time = datetime.now().strftime("%H%M%S")
file_name = ' '.join(tts_sentence.split(" ")[:10])
# if multiple sentences in one line -> split them
tts_sentence = split_into_sentences(tts_sentence)
# if sentence was split in sub-sentences -> iterate over them
for sentence in tts_sentence:
# synthesize voice
_, _, _, wav = tts(model,
vocoder,
C,
None,
sentence,
ap,
ap_vocoder,
use_cuda,
batched_vocoder,
speaker_id=speaker_id,
style_input=style_input,
figures=False)
# join sub-sentences back together and add a filler between them
wav_list += list(wav)
wav_list += [0] * 10000
wav = np.array(wav_list)
# finalize filename
file_name = "_".join([str(current_time), file_name])
file_name = file_name.translate(
str.maketrans('', '', string.punctuation.replace('_', ''))) + '.wav'
if out_path == "":
out_dir = str(Path(project, 'output', current_date, speaker_name))
out_path = os.path.join(out_dir, file_name)
else:
out_dir = os.path.dirname(out_path)
# create output directory if it doesn't exist
if not os.path.isdir(out_dir):
os.makedirs(out_dir, exist_ok=True)
# save generated wav to disk
ap.save_wav(wav, out_path)
end_time = time.time()
print(" > Run-time: {}".format(end_time - start_time))
print(" > Saving output to {}\n".format(out_path))
args = parser.parse_args()
if args.vocoder_path != "":
assert args.use_cuda, " [!] Enable cuda for vocoder."
from WaveRNN.models.wavernn import Model as VocoderModel
# load the config
C = load_config(args.config_path)
C.forward_attn_mask = True
# load the audio processor
ap = AudioProcessor(**C.audio)
# if the vocabulary was passed, replace the default
if 'text' in C.keys():
symbols, phonemes = make_symbols(**C.text)
# load speakers
if args.speakers_json != '':
speakers = json.load(open(args.speakers_json, 'r'))
num_speakers = len(speakers)
else:
num_speakers = 0
# load the model
num_chars = len(phonemes) if C.use_phonemes else len(symbols)
model = setup_model(num_chars, num_speakers, C)
cp = torch.load(args.model_path)
model.load_state_dict(cp['model'])
model.eval()
if args.use_cuda:
