class Synthesizer(object):
def __init__(self, config):
self.wavernn = None
self.config = config
self.use_cuda = config.use_cuda
if self.use_cuda:
assert torch.cuda.is_available(), "CUDA is not availabe on this machine."
self.load_tts(self.config.tts_path, self.config.tts_file, self.config.tts_config, config.use_cuda)
if self.config.wavernn_lib_path:
self.load_wavernn(config.wavernn_lib_path, config.wavernn_path, config.wavernn_file, config.wavernn_config, config.use_cuda)
def load_tts(self, model_path, model_file, model_config, use_cuda):
tts_config = os.path.join(model_path, model_config)
self.model_file = os.path.join(model_path, model_file)
print(" > Loading TTS model ...")
print(" | > model config: ", tts_config)
print(" | > model file: ", model_file)
self.tts_config = load_config(tts_config)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(**self.tts_config.audio)
if self.use_phonemes:
self.input_size = len(phonemes)
self.input_adapter = lambda sen: phoneme_to_sequence(sen, [self.tts_config.text_cleaner], self.tts_config.phoneme_language, self.tts_config.enable_eos_bos_chars)
else:
self.input_size = len(symbols)
self.input_adapter = lambda sen: text_to_sequence(sen, [self.tts_config.text_cleaner])
self.tts_model = setup_model(self.input_size, self.tts_config)
# load model state
if use_cuda:
cp = torch.load(self.model_file)
else:
cp = torch.load(self.model_file, map_location=lambda storage, loc: storage)
# load the model
self.tts_model.load_state_dict(cp['model'])
if use_cuda:
self.tts_model.cuda()
self.tts_model.eval()
self.tts_model.decoder.max_decoder_steps = 3000
def load_wavernn(self, lib_path, model_path, model_file, model_config, use_cuda):
sys.path.append(lib_path) # set this if TTS is not installed globally
from WaveRNN.models.wavernn import Model
wavernn_config = os.path.join(model_path, model_config)
model_file = os.path.join(model_path, model_file)
print(" > Loading WaveRNN model ...")
print(" | > model config: ", wavernn_config)
print(" | > model file: ", model_file)
self.wavernn_config = load_config(wavernn_config)
self.wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode=self.wavernn_config.mode,
pad=2,
upsample_factors=self.wavernn_config.upsample_factors, # set this depending on dataset
feat_dims=80,
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=self.ap.hop_length,
sample_rate=self.ap.sample_rate,
).cuda()
check = torch.load(model_file)
self.wavernn.load_state_dict(check['model'])
if use_cuda:
self.wavernn.cuda()
self.wavernn.eval()
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
def split_into_sentences(self, text):
text = " " + text + " "
text = text.replace("\n"," ")
text = re.sub(prefixes,"\\1<prd>",text)
text = re.sub(websites,"<prd>\\1",text)
if "Ph.D" in text: text = text.replace("Ph.D.","Ph<prd>D<prd>")
text = re.sub("\s" + alphabets + "[.] "," \\1<prd> ",text)
text = re.sub(acronyms+" "+starters,"\\1<stop> \\2",text)
text = re.sub(alphabets + "[.]" + alphabets + "[.]" + alphabets + "[.]","\\1<prd>\\2<prd>\\3<prd>",text)
text = re.sub(alphabets + "[.]" + alphabets + "[.]","\\1<prd>\\2<prd>",text)
text = re.sub(" "+suffixes+"[.] "+starters," \\1<stop> \\2",text)
text = re.sub(" "+suffixes+"[.]"," \\1<prd>",text)
text = re.sub(" " + alphabets + "[.]"," \\1<prd>",text)
if "”" in text: text = text.replace(".”","”.")
if "\"" in text: text = text.replace(".\"","\".")
if "!" in text: text = text.replace("!\"","\"!")
if "?" in text: text = text.replace("?\"","\"?")
text = text.replace(".",".<stop>")
text = text.replace("?","?<stop>")
text = text.replace("!","!<stop>")
text = text.replace("<prd>",".")
sentences = text.split("<stop>")
sentences = sentences[:-1]
sentences = [s.strip() for s in sentences]
return sentences
def tts(self, text):
wavs = []
sens = self.split_into_sentences(text)
if len(sens) == 0:
sens = [text+'.']
for sen in sens:
if len(sen) < 3:
continue
sen = sen.strip()
print(sen)
class Synthesizer(object):
def __init__(self, config):
self.wavernn = None
self.config = config
self.use_cuda = config.use_cuda
if self.use_cuda:
assert torch.cuda.is_available(
), "CUDA is not availabe on this machine."
self.load_tts(self.config.tts_path, self.config.tts_file,
self.config.tts_config, config.use_cuda)
if self.config.wavernn_lib_path:
self.load_wavernn(config.wavernn_lib_path, config.wavernn_path,
config.wavernn_file, config.wavernn_config,
config.use_cuda)
def load_tts(self, model_path, model_file, model_config, use_cuda):
tts_config = os.path.join(model_path, model_config)
self.model_file = os.path.join(model_path, model_file)
print(" > Loading TTS model ...")
print(" | > model config: ", tts_config)
print(" | > model file: ", model_file)
self.tts_config = load_config(tts_config)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(**self.tts_config.audio)
if self.use_phonemes:
self.input_size = len(phonemes)
self.input_adapter = lambda sen: phoneme_to_sequence(
sen, [self.tts_config.text_cleaner], self.tts_config.
phoneme_language, self.tts_config.enable_eos_bos_chars)
else:
self.input_size = len(symbols)
self.input_adapter = lambda sen: text_to_sequence(
sen, [self.tts_config.text_cleaner])
# load speakers
if self.config.tts_speakers is not None:
self.tts_speakers = load_speaker_mapping(
os.path.join(model_path, self.config.tts_speakers))
num_speakers = len(self.tts_speakers)
else:
num_speakers = 0
self.tts_model = setup_model(self.input_size,
num_speakers=num_speakers,
c=self.tts_config)
# load model state
if use_cuda:
cp = torch.load(self.model_file)
else:
cp = torch.load(self.model_file,
map_location=lambda storage, loc: storage)
# load the model
self.tts_model.load_state_dict(cp['model'])
if use_cuda:
self.tts_model.cuda()
self.tts_model.eval()
self.tts_model.decoder.max_decoder_steps = 3000
def load_wavernn(self, lib_path, model_path, model_file, model_config,
use_cuda):
# TODO: set a function in wavernn code base for model setup and call it here.
sys.path.append(lib_path) # set this if TTS is not installed globally
from WaveRNN.models.wavernn import Model
wavernn_config = os.path.join(model_path, model_config)
model_file = os.path.join(model_path, model_file)
print(" > Loading WaveRNN model ...")
print(" | > model config: ", wavernn_config)
print(" | > model file: ", model_file)
self.wavernn_config = load_config(wavernn_config)
self.wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode=self.wavernn_config.mode,
mulaw=self.wavernn_config.mulaw,
pad=self.wavernn_config.pad,
use_aux_net=self.wavernn_config.use_aux_net,
use_upsample_net=self.wavernn_config.use_upsample_net,
upsample_factors=self.wavernn_config.upsample_factors,
feat_dims=80,
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=self.ap.hop_length,
sample_rate=self.ap.sample_rate,
).cuda()
check = torch.load(model_file)
self.wavernn.load_state_dict(check['model'])
if use_cuda:
self.wavernn.cuda()
self.wavernn.eval()
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
def split_into_sentences(self, text):
text = " " + text + " "
text = text.replace("\n", " ")
text = re.sub(prefixes, "\\1<prd>", text)
text = re.sub(websites, "<prd>\\1", text)
if "Ph.D" in text:
text = text.replace("Ph.D.", "Ph<prd>D<prd>")
text = re.sub(r"\s" + alphabets + "[.] ", " \\1<prd> ", text)
text = re.sub(acronyms + " " + starters, "\\1<stop> \\2", text)
text = re.sub(
alphabets + "[.]" + alphabets + "[.]" + alphabets + "[.]",
"\\1<prd>\\2<prd>\\3<prd>", text)
text = re.sub(alphabets + "[.]" + alphabets + "[.]",
"\\1<prd>\\2<prd>", text)
text = re.sub(" " + suffixes + "[.] " + starters, " \\1<stop> \\2",
text)
text = re.sub(" " + suffixes + "[.]", " \\1<prd>", text)
text = re.sub(" " + alphabets + "[.]", " \\1<prd>", text)
if "”" in text:
text = text.replace(".”", "”.")
if "\"" in text:
text = text.replace(".\"", "\".")
if "!" in text:
text = text.replace("!\"", "\"!")
if "?" in text:
text = text.replace("?\"", "\"?")
text = text.replace(".", ".<stop>")
text = text.replace("?", "?<stop>")
text = text.replace("!", "!<stop>")
text = text.replace("<prd>", ".")
sentences = text.split("<stop>")
sentences = sentences[:-1]
sentences = [s.strip() for s in sentences]
return sentences
def tts(self, text):
wavs = []
sens = self.split_into_sentences(text)
if not sens:
sens = [text + '.']
for sen in sens:
if len(sen) < 3:
continue
sen = sen.strip()
print(sen)
seq = np.array(self.input_adapter(sen))
text_hat = sequence_to_phoneme(seq)
print(text_hat)
chars_var = torch.from_numpy(seq).unsqueeze(0).long()
if self.use_cuda:
chars_var = chars_var.cuda()
decoder_out, postnet_out, alignments, stop_tokens = self.tts_model.inference(
chars_var)
postnet_out = postnet_out[0].data.cpu().numpy()
if self.tts_config.model == "Tacotron":
wav = self.ap.inv_spectrogram(postnet_out.T)
elif self.tts_config.model == "Tacotron2":
if self.wavernn:
wav = self.wavernn.generate(
torch.FloatTensor(postnet_out.T).unsqueeze(0).cuda(),
batched=self.config.is_wavernn_batched,
target=11000,
overlap=550)
else:
wav = self.ap.inv_mel_spectrogram(postnet_out.T)
wavs += list(wav)
wavs += [0] * 10000
out = io.BytesIO()
self.save_wav(wavs, out)
return out
class Synthesizer(object):
def __init__(self, config):
self.wavernn = None
self.pwgan = None
self.config = config
self.use_cuda = self.config.use_cuda
if self.use_cuda:
assert torch.cuda.is_available(
), "CUDA is not availabe on this machine."
self.load_tts(self.config.tts_checkpoint, self.config.tts_config,
self.config.use_cuda)
if self.config.vocoder_checkpoint:
self.load_vocoder(self.config.vocoder_checkpoint,
self.config.vocoder_config, self.config.use_cuda)
if self.config.wavernn_lib_path:
self.load_wavernn(self.config.wavernn_lib_path,
self.config.wavernn_file,
self.config.wavernn_config, self.config.use_cuda)
if self.config.pwgan_file:
self.load_pwgan(self.config.pwgan_lib_path, self.config.pwgan_file,
self.config.pwgan_config, self.config.use_cuda)
def load_tts(self, tts_checkpoint, tts_config, use_cuda):
# pylint: disable=global-statement
global symbols, phonemes
print(" > Loading TTS model ...")
print(" | > model config: ", tts_config)
print(" | > checkpoint file: ", tts_checkpoint)
self.tts_config = load_config(tts_config)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(**self.tts_config.audio)
if 'characters' in self.tts_config.keys():
symbols, phonemes = make_symbols(**self.tts_config.characters)
if self.use_phonemes:
self.input_size = len(phonemes)
else:
self.input_size = len(symbols)
# TODO: fix this for multi-speaker model - load speakers
if self.config.tts_speakers is not None:
self.tts_speakers = load_speaker_mapping(self.config.tts_speakers)
num_speakers = len(self.tts_speakers)
else:
num_speakers = 0
self.tts_model = setup_model(self.input_size,
num_speakers=num_speakers,
c=self.tts_config)
# load model state
cp = torch.load(tts_checkpoint, map_location=torch.device('cpu'))
# load the model
self.tts_model.load_state_dict(cp['model'])
if use_cuda:
self.tts_model.cuda()
self.tts_model.eval()
self.tts_model.decoder.max_decoder_steps = 3000
if 'r' in cp:
self.tts_model.decoder.set_r(cp['r'])
print(f" > model reduction factor: {cp['r']}")
def load_vocoder(self, model_file, model_config, use_cuda):
self.vocoder_config = load_config(model_config)
self.vocoder_model = setup_generator(self.vocoder_config)
self.vocoder_model.load_state_dict(
torch.load(model_file, map_location="cpu")["model"])
self.vocoder_model.remove_weight_norm()
self.vocoder_model.inference_padding = 0
self.vocoder_config = load_config(model_config)
if use_cuda:
self.vocoder_model.cuda()
self.vocoder_model.eval()
def load_wavernn(self, lib_path, model_file, model_config, use_cuda):
# TODO: set a function in wavernn code base for model setup and call it here.
sys.path.append(
lib_path) # set this if WaveRNN is not installed globally
#pylint: disable=import-outside-toplevel
from WaveRNN.models.wavernn import Model
print(" > Loading WaveRNN model ...")
print(" | > model config: ", model_config)
print(" | > model file: ", model_file)
self.wavernn_config = load_config(model_config)
# This is the default architecture we use for our models.
# You might need to update it
self.wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode=self.wavernn_config.mode,
mulaw=self.wavernn_config.mulaw,
pad=self.wavernn_config.pad,
use_aux_net=self.wavernn_config.use_aux_net,
use_upsample_net=self.wavernn_config.use_upsample_net,
upsample_factors=self.wavernn_config.upsample_factors,
feat_dims=80,
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=self.ap.hop_length,
sample_rate=self.ap.sample_rate,
).cuda()
check = torch.load(model_file, map_location="cpu")
self.wavernn.load_state_dict(check['model'])
if use_cuda:
self.wavernn.cuda()
self.wavernn.eval()
def load_pwgan(self, lib_path, model_file, model_config, use_cuda):
if lib_path:
# set this if ParallelWaveGAN is not installed globally
sys.path.append(lib_path)
try:
#pylint: disable=import-outside-toplevel
from parallel_wavegan.models import ParallelWaveGANGenerator
except ImportError as e:
raise RuntimeError(
f"cannot import parallel-wavegan, either install it or set its directory using the --pwgan_lib_path command line argument: {e}"
)
print(" > Loading PWGAN model ...")
print(" | > model config: ", model_config)
print(" | > model file: ", model_file)
with open(model_config) as f:
self.pwgan_config = yaml.load(f, Loader=yaml.Loader)
self.pwgan = ParallelWaveGANGenerator(
**self.pwgan_config["generator_params"])
self.pwgan.load_state_dict(
torch.load(model_file, map_location="cpu")["model"]["generator"])
self.pwgan.remove_weight_norm()
if use_cuda:
self.pwgan.cuda()
self.pwgan.eval()
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
@staticmethod
def split_into_sentences(text):
text = " " + text + " <stop>"
text = text.replace("\n", " ")
text = re.sub(prefixes, "\\1<prd>", text)
text = re.sub(websites, "<prd>\\1", text)
if "Ph.D" in text:
text = text.replace("Ph.D.", "Ph<prd>D<prd>")
text = re.sub(r"\s" + alphabets + "[.] ", " \\1<prd> ", text)
text = re.sub(acronyms + " " + starters, "\\1<stop> \\2", text)
text = re.sub(
alphabets + "[.]" + alphabets + "[.]" + alphabets + "[.]",
"\\1<prd>\\2<prd>\\3<prd>", text)
text = re.sub(alphabets + "[.]" + alphabets + "[.]",
"\\1<prd>\\2<prd>", text)
text = re.sub(" " + suffixes + "[.] " + starters, " \\1<stop> \\2",
text)
text = re.sub(" " + suffixes + "[.]", " \\1<prd>", text)
text = re.sub(" " + alphabets + "[.]", " \\1<prd>", text)
if "”" in text:
text = text.replace(".”", "”.")
if "\"" in text:
text = text.replace(".\"", "\".")
if "!" in text:
text = text.replace("!\"", "\"!")
if "?" in text:
text = text.replace("?\"", "\"?")
text = text.replace(".", ".<stop>")
text = text.replace("?", "?<stop>")
text = text.replace("!", "!<stop>")
text = text.replace("<prd>", ".")
sentences = text.split("<stop>")
sentences = sentences[:-1]
sentences = list(filter(
None, [s.strip() for s in sentences])) # remove empty sentences
return sentences
def tts(self, text, speaker_id=None):
start_time = time.time()
wavs = []
sens = self.split_into_sentences(text)
print(sens)
speaker_id = id_to_torch(speaker_id)
if speaker_id is not None and self.use_cuda:
speaker_id = speaker_id.cuda()
for sen in sens:
# preprocess the given text
inputs = text_to_seqvec(sen, self.tts_config)
inputs = numpy_to_torch(inputs, torch.long, cuda=self.use_cuda)
inputs = inputs.unsqueeze(0)
# synthesize voice
decoder_output, postnet_output, alignments, stop_tokens = run_model_torch(
self.tts_model, inputs, self.tts_config, False, speaker_id,
None)
# convert outputs to numpy
if self.vocoder_model:
vocoder_input = postnet_output[0].transpose(0, 1).unsqueeze(0)
wav = self.vocoder_model.inference(vocoder_input)
if self.use_cuda:
wav = wav.cpu().numpy()
else:
wav = wav.numpy()
wav = wav.flatten()
elif self.wavernn:
vocoder_input = None
if self.tts_config.model == "Tacotron":
vocoder_input = torch.FloatTensor(
self.ap.out_linear_to_mel(
linear_spec=postnet_output.T).T).T.unsqueeze(0)
else:
vocoder_input = postnet_output[0].transpose(0,
1).unsqueeze(0)
if self.use_cuda:
vocoder_input.cuda()
wav = self.wavernn.generate(
vocoder_input,
batched=self.config.is_wavernn_batched,
target=11000,
overlap=550)
# trim silence
wav = trim_silence(wav, self.ap)
wavs += list(wav)
wavs += [0] * 10000
out = io.BytesIO()
self.save_wav(wavs, out)
# compute stats
process_time = time.time() - start_time
audio_time = len(wavs) / self.tts_config.audio['sample_rate']
print(f" > Processing time: {process_time}")
print(f" > Real-time factor: {process_time / audio_time}")
return out
mode=VC.mode,
mulaw=VC.mulaw,
pad=VC.pad,
upsample_factors=VC.upsample_factors,
feat_dims=VC.audio["num_mels"],
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=ap.hop_length,
sample_rate=ap.sample_rate,
use_aux_net=True,
use_upsample_net=True)
check = torch.load(args.vocoder_path)
vocoder_model.load_state_dict(check['model'])
vocoder_model.eval()
if args.use_cuda:
vocoder_model.cuda()
else:
vocoder_model = None
VC = None
ap_vocoder = None
# synthesize voice
print(" > Text: {}".format(args.text))
_, _, _, wav = tts(model,
vocoder_model,
C,
VC,
args.text,
ap,
wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode=VOCODER_CONFIG.mode,
mulaw=VOCODER_CONFIG.mulaw,
pad=VOCODER_CONFIG.pad,
use_aux_net=VOCODER_CONFIG.use_aux_net,
use_upsample_net=VOCODER_CONFIG.use_upsample_net,
upsample_factors=VOCODER_CONFIG.upsample_factors,
feat_dims=80,
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=ap2.hop_length,
sample_rate=ap2.sample_rate,
)
check = torch.load(VOCODER_MODEL_PATH,
map_location=torch.device('cuda' if use_cuda else 'cpu'))
wavernn.load_state_dict(check['model'])
if use_cuda:
wavernn.cuda()
wavernn.eval()
model.decoder.max_decoder_steps = 500
br = 50
stay = True
while (stay):
synthesise_text(input('Type a sentence to be synthesised > '))
stay = not (input('Type exit to stop > ') == 'exit')
class Synthesizer(object):
def __init__(self, config):
self.wavernn = None
self.vocoder_model = None
self.config = config
print(config)
self.seg = self.get_segmenter("en")
self.use_cuda = self.config.use_cuda
if self.use_cuda:
assert torch.cuda.is_available(
), "CUDA is not availabe on this machine."
self.load_tts(self.config.tts_checkpoint, self.config.tts_config,
self.config.use_cuda)
if self.config.vocoder_checkpoint:
self.load_vocoder(self.config.vocoder_checkpoint,
self.config.vocoder_config, self.config.use_cuda)
if self.config.wavernn_lib_path:
self.load_wavernn(self.config.wavernn_lib_path,
self.config.wavernn_checkpoint,
self.config.wavernn_config, self.config.use_cuda)
@staticmethod
def get_segmenter(lang):
return pysbd.Segmenter(language=lang, clean=True)
def load_tts(self, tts_checkpoint, tts_config, use_cuda):
# pylint: disable=global-statement
global symbols, phonemes
print(" > Loading TTS model ...")
print(" | > model config: ", tts_config)
print(" | > checkpoint file: ", tts_checkpoint)
self.tts_config = load_config(tts_config)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(**self.tts_config.audio)
if 'characters' in self.tts_config.keys():
symbols, phonemes = make_symbols(**self.tts_config.characters)
if self.use_phonemes:
self.input_size = len(phonemes)
else:
self.input_size = len(symbols)
# TODO: fix this for multi-speaker model - load speakers
if self.config.tts_speakers is not None:
self.tts_speakers = load_speaker_mapping(self.config.tts_speakers)
num_speakers = len(self.tts_speakers)
else:
num_speakers = 0
self.tts_model = setup_model(self.input_size,
num_speakers=num_speakers,
c=self.tts_config)
# load model state
cp = torch.load(tts_checkpoint, map_location=torch.device('cpu'))
# load the model
self.tts_model.load_state_dict(cp['model'])
if use_cuda:
self.tts_model.cuda()
self.tts_model.eval()
self.tts_model.decoder.max_decoder_steps = 3000
if 'r' in cp:
self.tts_model.decoder.set_r(cp['r'])
print(f" > model reduction factor: {cp['r']}")
def load_vocoder(self, model_file, model_config, use_cuda):
self.vocoder_config = load_config(model_config)
self.vocoder_model = setup_generator(self.vocoder_config)
self.vocoder_model.load_state_dict(
torch.load(model_file, map_location="cpu")["model"])
self.vocoder_model.remove_weight_norm()
self.vocoder_model.inference_padding = 0
self.vocoder_config = load_config(model_config)
if use_cuda:
self.vocoder_model.cuda()
self.vocoder_model.eval()
def load_wavernn(self, lib_path, model_file, model_config, use_cuda):
# TODO: set a function in wavernn code base for model setup and call it here.
sys.path.append(
lib_path) # set this if WaveRNN is not installed globally
#pylint: disable=import-outside-toplevel
from WaveRNN.models.wavernn import Model
print(" > Loading WaveRNN model ...")
print(" | > model config: ", model_config)
print(" | > model file: ", model_file)
self.wavernn_config = load_config(model_config)
# This is the default architecture we use for our models.
# You might need to update it
self.wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode=self.wavernn_config.mode,
mulaw=self.wavernn_config.mulaw,
pad=self.wavernn_config.pad,
use_aux_net=self.wavernn_config.use_aux_net,
use_upsample_net=self.wavernn_config.use_upsample_net,
upsample_factors=self.wavernn_config.upsample_factors,
feat_dims=80,
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=self.ap.hop_length,
sample_rate=self.ap.sample_rate,
).cuda()
check = torch.load(model_file, map_location="cpu")
self.wavernn.load_state_dict(check['model'])
if use_cuda:
self.wavernn.cuda()
self.wavernn.eval()
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
def split_into_sentences(self, text):
return self.seg.segment(text)
def tts(self, text, speaker_id=None):
start_time = time.time()
wavs = []
sens = self.split_into_sentences(text)
print(sens)
speaker_id = id_to_torch(speaker_id)
if speaker_id is not None and self.use_cuda:
speaker_id = speaker_id.cuda()
for sen in sens:
# preprocess the given text
inputs = text_to_seqvec(sen, self.tts_config)
inputs = numpy_to_torch(inputs, torch.long, cuda=self.use_cuda)
inputs = inputs.unsqueeze(0)
# synthesize voice
_, postnet_output, _, _ = run_model_torch(self.tts_model, inputs,
self.tts_config, False,
speaker_id, None)
if self.vocoder_model:
# use native vocoder model
vocoder_input = postnet_output[0].transpose(0, 1).unsqueeze(0)
wav = self.vocoder_model.inference(vocoder_input)
if self.use_cuda:
wav = wav.cpu().numpy()
else:
wav = wav.numpy()
wav = wav.flatten()
elif self.wavernn:
# use 3rd paty wavernn
vocoder_input = None
if self.tts_config.model == "Tacotron":
vocoder_input = torch.FloatTensor(
self.ap.out_linear_to_mel(
linear_spec=postnet_output.T).T).T.unsqueeze(0)
else:
vocoder_input = postnet_output[0].transpose(0,
1).unsqueeze(0)
if self.use_cuda:
vocoder_input.cuda()
wav = self.wavernn.generate(
vocoder_input,
batched=self.config.is_wavernn_batched,
target=11000,
overlap=550)
else:
# use GL
if self.use_cuda:
postnet_output = postnet_output[0].cpu()
else:
postnet_output = postnet_output[0]
postnet_output = postnet_output.numpy()
wav = inv_spectrogram(postnet_output, self.ap, self.tts_config)
# trim silence
wav = trim_silence(wav, self.ap)
wavs += list(wav)
wavs += [0] * 10000
out = io.BytesIO()
self.save_wav(wavs, out)
# compute stats
process_time = time.time() - start_time
audio_time = len(wavs) / self.tts_config.audio['sample_rate']
print(f" > Processing time: {process_time}")
print(f" > Real-time factor: {process_time / audio_time}")
return out
class Synthesizer(object):
def __init__(self, config):
self.wavernn = None
self.config = config
self.use_cuda = self.config.use_cuda
if self.use_cuda:
assert torch.cuda.is_available(), "CUDA is not availabe on this machine."
tts_config_path = resource_path(self.config.tts_config)
tts_checkpoint_path = resource_path(self.config.tts_checkpoint)
self.load_tts(tts_checkpoint_path, tts_config_path, self.config.use_cuda)
if self.config.wavernn_lib_path:
wavernn_lib_path = resource_path(self.config.wavernn_lib_path)
wavernn_path = resource_path(self.config.wavernn_path)
self.load_wavernn(wavernn_lib_path, wavernn_path, wavernn_file, wavernn_config, self.config.use_cuda)
def load_tts(self, tts_checkpoint, tts_config, use_cuda):
print("\t[-] Loading TTS model ...")
print("\t | > model config: ", tts_config)
print("\t | > checkpoint file: ", tts_checkpoint)
self.tts_config = load_config(tts_config)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(**self.tts_config.audio)
if self.use_phonemes:
self.input_size = len(phonemes)
else:
self.input_size = len(symbols)
if self.config.tts_speakers is not None:
self.tts_speakers = load_speaker_mapping(os.path.join(model_path, self.config.tts_speakers))
num_speakers = len(self.tts_speakers)
else:
num_speakers = 0
self.tts_model = setup_model(self.input_size, num_speakers=num_speakers, c=self.tts_config)
cp = torch.load(tts_checkpoint, map_location=torch.device('cpu'))
self.tts_model.load_state_dict(cp['model'])
if use_cuda:
self.tts_model.cuda()
self.tts_model.eval()
self.tts_model.decoder.max_decoder_steps = 3000
if 'r' in cp:
self.tts_model.decoder.set_r(cp['r'])
def load_wavernn(self, lib_path, model_path, model_file, model_config, use_cuda):
sys.path.append(lib_path)
from WaveRNN.models.wavernn import Model
wavernn_config = os.path.join(model_path, model_config)
model_file = os.path.join(model_path, model_file)
print("\t[-] Loading WaveRNN model ...")
print("\t | > model config: ", wavernn_config)
print("\t | > model file: ", model_file)
self.wavernn_config = load_config(wavernn_config)
self.wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode=self.wavernn_config.mode,
mulaw=self.wavernn_config.mulaw,
pad=self.wavernn_config.pad,
use_aux_net=self.wavernn_config.use_aux_net,
use_upsample_net = self.wavernn_config.use_upsample_net,
upsample_factors=self.wavernn_config.upsample_factors,
feat_dims=80,
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=self.ap.hop_length,
sample_rate=self.ap.sample_rate,
).cuda()
check = torch.load(model_file)
self.wavernn.load_state_dict(check['model'])
if use_cuda:
self.wavernn.cuda()
self.wavernn.eval()
def save_wav(self, wav, path):
wav = np.array(wav)
self.ap.save_wav(wav, path)
def split_into_sentences(self, text):
text = " " + text + " "
text = text.replace("\n", " ")
text = re.sub(prefixes, "\\1<prd>", text)
text = re.sub(websites, "<prd>\\1", text)
if "Ph.D" in text:
text = text.replace("Ph.D.", "Ph<prd>D<prd>")
text = re.sub(r"\s" + alphabets + "[.] ", " \\1<prd> ", text)
text = re.sub(acronyms+" "+starters, "\\1<stop> \\2", text)
text = re.sub(alphabets + "[.]" + alphabets + "[.]" + alphabets + "[.]", "\\1<prd>\\2<prd>\\3<prd>", text)
text = re.sub(alphabets + "[.]" + alphabets + "[.]", "\\1<prd>\\2<prd>", text)
text = re.sub(" "+suffixes+"[.] "+starters, " \\1<stop> \\2", text)
text = re.sub(" "+suffixes+"[.]", " \\1<prd>", text)
text = re.sub(" " + alphabets + "[.]", " \\1<prd>", text)
if "”" in text:
text = text.replace(".”", "”.")
if "\"" in text:
text = text.replace(".\"", "\".")
if "!" in text:
text = text.replace("!\"", "\"!")
if "?" in text:
text = text.replace("?\"", "\"?")
text = text.replace(".", ".<stop>")
text = text.replace("?", "?<stop>")
text = text.replace("!", "!<stop>")
text = text.replace("<prd>", ".")
sentences = text.split("<stop>")
sentences = sentences[:-1]
sentences = [s.strip() for s in sentences]
return sentences
def tts(self, text):
text = f"{text}." # Adding a fullstop to the end
wavs = []
sens = self.split_into_sentences(text)
if not sens:
sens = [text+'.']
for sen in sens:
inputs = text_to_seqvec(sen, self.tts_config, self.use_cuda)
decoder_output, postnet_output, alignments, _ = run_model(
self.tts_model, inputs, self.tts_config, False, None, None)
postnet_output, decoder_output, _ = parse_outputs(postnet_output, decoder_output, alignments)
if self.wavernn:
postnet_output = postnet_output[0].data.cpu().numpy()
wav = self.wavernn.generate(
torch.FloatTensor(postnet_output.T).unsqueeze(0).cuda(),
batched=self.config.is_wavernn_batched,
target=11000,
overlap=550)
else:
wav = inv_spectrogram(
postnet_output,
self.ap,
self.tts_config)
wav = trim_silence(wav, self.ap)
wavs += list(wav)
wavs += [0] * 10000
out = io.BytesIO()
self.save_wav(wavs, out)
return out
class Synthesizer(object):
def __init__(self, config):
self.wavernn = None
self.config = config
self.use_cuda = self.config.use_cuda
if self.use_cuda:
assert torch.cuda.is_available(
), "CUDA is not availabe on this machine."
self.load_tts(self.config.tts_path, self.config.tts_file,
self.config.tts_config, config.use_cuda)
if self.config.wavernn_lib_path:
self.load_wavernn(config.wavernn_lib_path, config.wavernn_path,
config.wavernn_file, config.wavernn_config,
config.use_cuda)
def load_tts(self, model_path, model_file, model_config, use_cuda):
tts_config = os.path.join(model_path, model_config)
self.model_file = os.path.join(model_path, model_file)
self.tts_config = load_config(tts_config)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(**self.tts_config.audio)
if self.use_phonemes:
self.input_size = len(phonemes)
else:
self.input_size = len(symbols)
# load speakers
if self.config.tts_speakers is not None:
self.tts_speakers = load_speaker_mapping(
os.path.join(model_path, self.config.tts_speakers))
num_speakers = len(self.tts_speakers)
else:
num_speakers = 0
self.tts_model = setup_model(self.input_size,
num_speakers=num_speakers,
c=self.tts_config)
# load model state
cp = torch.load(self.model_file, map_location=torch.device('cpu'))
# load the model
self.tts_model.load_state_dict(cp['model'])
if use_cuda:
self.tts_model.cuda()
self.tts_model.eval()
self.tts_model.decoder.max_decoder_steps = 3000
if 'r' in cp and self.tts_config.model in ["Tacotron", "TacotronGST"]:
self.tts_model.decoder.set_r(cp['r'])
def load_wavernn(self, lib_path, model_path, model_file, model_config,
use_cuda):
# TODO: set a function in wavernn code base for model setup and call it here.
sys.path.append(lib_path) # set this if TTS is not installed globally
from WaveRNN.models.wavernn import Model
wavernn_config = os.path.join(model_path, model_config)
model_file = os.path.join(model_path, model_file)
self.wavernn_config = load_config(wavernn_config)
self.wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode=self.wavernn_config.mode,
mulaw=self.wavernn_config.mulaw,
pad=self.wavernn_config.pad,
use_aux_net=self.wavernn_config.use_aux_net,
use_upsample_net=self.wavernn_config.use_upsample_net,
upsample_factors=self.wavernn_config.upsample_factors,
feat_dims=80,
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=self.ap.hop_length,
sample_rate=self.ap.sample_rate,
).cuda()
check = torch.load(model_file, map_location=torch.device('cpu'))
self.wavernn.load_state_dict(check['model'])
if use_cuda:
self.wavernn.cuda()
self.wavernn.eval()
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
def split_into_sentences(self, text):
text = " " + text + " "
text = text.replace("\n", " ")
text = re.sub(prefixes, "\\1<prd>", text)
text = re.sub(websites, "<prd>\\1", text)
if "Ph.D" in text:
text = text.replace("Ph.D.", "Ph<prd>D<prd>")
text = re.sub(r"\s" + alphabets + "[.] ", " \\1<prd> ", text)
text = re.sub(acronyms + " " + starters, "\\1<stop> \\2", text)
text = re.sub(
alphabets + "[.]" + alphabets + "[.]" + alphabets + "[.]",
"\\1<prd>\\2<prd>\\3<prd>", text)
text = re.sub(alphabets + "[.]" + alphabets + "[.]",
"\\1<prd>\\2<prd>", text)
text = re.sub(" " + suffixes + "[.] " + starters, " \\1<stop> \\2",
text)
text = re.sub(" " + suffixes + "[.]", " \\1<prd>", text)
text = re.sub(" " + alphabets + "[.]", " \\1<prd>", text)
if "”" in text:
text = text.replace(".”", "”.")
if "\"" in text:
text = text.replace(".\"", "\".")
if "!" in text:
text = text.replace("!\"", "\"!")
if "?" in text:
text = text.replace("?\"", "\"?")
text = text.replace(".", ".<stop>")
text = text.replace("?", "?<stop>")
text = text.replace("!", "!<stop>")
text = text.replace("<prd>", ".")
sentences = text.split("<stop>")
sentences = sentences[:-1]
sentences = [s.strip() for s in sentences]
return sentences
def tts(self, sens):
wavs = []
#sens = self.split_into_sentences(text)
for sen in sens:
# preprocess the given text
inputs = text_to_seqvec(sen, self.tts_config, self.use_cuda)
# synthesize voice
decoder_output, postnet_output, alignments, _ = run_model(
self.tts_model, inputs, self.tts_config, False, None, None)
# convert outputs to numpy
postnet_output, decoder_output, _ = parse_outputs(
postnet_output, decoder_output, alignments)
if self.wavernn:
postnet_output = postnet_output[0].data.cpu().numpy()
wav = self.wavernn.generate(
torch.FloatTensor(postnet_output.T).unsqueeze(0).cuda(),
batched=self.config.is_wavernn_batched,
target=11000,
overlap=550)
else:
wav = inv_spectrogram(postnet_output, self.ap, self.tts_config)
# trim silence
wav = trim_silence(wav, self.ap)
wavs += list(wav)
wavs += [0] * 1000
out = io.BytesIO()
self.save_wav(wavs, out)
return out
class Synthesizer(object):
def __init__(self, config):
self.wavernn = None
self.pwgan = None
self.config = config
self.use_cuda = self.config.use_cuda
if self.use_cuda:
assert torch.cuda.is_available(
), "CUDA is not availabe on this machine."
self.load_tts(self.config.tts_checkpoint, self.config.tts_config,
self.config.use_cuda)
if self.config.wavernn_lib_path:
self.load_wavernn(self.config.wavernn_lib_path,
self.config.wavernn_file,
self.config.wavernn_config, self.config.use_cuda)
if self.config.pwgan_file:
self.load_pwgan(self.config.pwgan_file, self.config.pwgan_config,
self.config.use_cuda)
def load_tts(self, tts_checkpoint, tts_config, use_cuda):
print(" > Loading TTS model ...")
print(" | > model config: ", tts_config)
print(" | > checkpoint file: ", tts_checkpoint)
self.tts_config = load_config(tts_config)
self.use_phonemes = self.tts_config.use_phonemes
self.ap = AudioProcessor(**self.tts_config.audio)
if self.use_phonemes:
self.input_size = len(phonemes)
else:
self.input_size = len(symbols)
# TODO: fix this for multi-speaker model - load speakers
if self.config.tts_speakers is not None:
self.tts_speakers = load_speaker_mapping(self.config.tts_speakers)
num_speakers = len(self.tts_speakers)
else:
num_speakers = 0
self.tts_model = setup_model(self.input_size,
num_speakers=num_speakers,
c=self.tts_config)
# load model state
cp = torch.load(tts_checkpoint, map_location=torch.device('cpu'))
# load the model
self.tts_model.load_state_dict(cp['model'])
if use_cuda:
self.tts_model.cuda()
self.tts_model.eval()
self.tts_model.decoder.max_decoder_steps = 3000
if 'r' in cp:
self.tts_model.decoder.set_r(cp['r'])
def load_wavernn(self, lib_path, model_file, model_config, use_cuda):
# TODO: set a function in wavernn code base for model setup and call it here.
sys.path.append(
lib_path) # set this if WaveRNN is not installed globally
#pylint: disable=import-outside-toplevel
from WaveRNN.models.wavernn import Model
print(" > Loading WaveRNN model ...")
print(" | > model config: ", model_config)
print(" | > model file: ", model_file)
self.wavernn_config = load_config(model_config)
# This is the default architecture we use for our models.
# You might need to update it
self.wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode=self.wavernn_config.mode,
mulaw=self.wavernn_config.mulaw,
pad=self.wavernn_config.pad,
use_aux_net=self.wavernn_config.use_aux_net,
use_upsample_net=self.wavernn_config.use_upsample_net,
upsample_factors=self.wavernn_config.upsample_factors,
feat_dims=80,
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=self.ap.hop_length,
sample_rate=self.ap.sample_rate,
).cuda()
check = torch.load(model_file)
self.wavernn.load_state_dict(check['model'], map_location="cpu")
if use_cuda:
self.wavernn.cuda()
self.wavernn.eval()
def load_pwgan(self, model_file, model_config, use_cuda):
#pylint: disable=import-outside-toplevel
from parallel_wavegan.models import ParallelWaveGANGenerator
from parallel_wavegan.utils.audio import AudioProcessor as AudioProcessorVocoder
print(" > Loading PWGAN model ...")
print(" | > model config: ", model_config)
print(" | > model file: ", model_file)
with open(model_config) as f:
self.pwgan_config = yaml.load(f, Loader=yaml.Loader)
self.pwgan = ParallelWaveGANGenerator(
**self.pwgan_config["generator_params"])
self.pwgan.load_state_dict(
torch.load(model_file, map_location="cpu")["model"]["generator"])
self.pwgan.remove_weight_norm()
self.pwgan_ap = AudioProcessorVocoder(**self.pwgan_config["audio"])
if use_cuda:
self.pwgan.cuda()
self.pwgan.eval()
def save_wav(self, wav, path):
# wav *= 32767 / max(1e-8, np.max(np.abs(wav)))
wav = np.array(wav)
self.ap.save_wav(wav, path)
def split_into_sentences(self, text):
text = " " + text + " <stop>"
text = text.replace("\n", " ")
text = re.sub(prefixes, "\\1<prd>", text)
text = re.sub(websites, "<prd>\\1", text)
if "Ph.D" in text:
text = text.replace("Ph.D.", "Ph<prd>D<prd>")
text = re.sub(r"\s" + alphabets + "[.] ", " \\1<prd> ", text)
text = re.sub(acronyms + " " + starters, "\\1<stop> \\2", text)
text = re.sub(
alphabets + "[.]" + alphabets + "[.]" + alphabets + "[.]",
"\\1<prd>\\2<prd>\\3<prd>", text)
text = re.sub(alphabets + "[.]" + alphabets + "[.]",
"\\1<prd>\\2<prd>", text)
text = re.sub(" " + suffixes + "[.] " + starters, " \\1<stop> \\2",
text)
text = re.sub(" " + suffixes + "[.]", " \\1<prd>", text)
text = re.sub(" " + alphabets + "[.]", " \\1<prd>", text)
if "”" in text:
text = text.replace(".”", "”.")
if "\"" in text:
text = text.replace(".\"", "\".")
if "!" in text:
text = text.replace("!\"", "\"!")
if "?" in text:
text = text.replace("?\"", "\"?")
text = text.replace(".", ".<stop>")
text = text.replace("?", "?<stop>")
text = text.replace("!", "!<stop>")
text = text.replace("<prd>", ".")
sentences = text.split("<stop>")
sentences = sentences[:-1]
sentences = list(filter(None, [s.strip() for s in sentences]))
return sentences
def tts(self, text):
wavs = []
sens = self.split_into_sentences(text)
print(sens)
if not sens:
sens = [text + '.']
for sen in sens:
# preprocess the given text
inputs = text_to_seqvec(sen, self.tts_config, self.use_cuda)
# synthesize voice
decoder_output, postnet_output, alignments, _ = run_model(
self.tts_model, inputs, self.tts_config, False, None, None)
# convert outputs to numpy
postnet_output, decoder_output, _ = parse_outputs(
postnet_output, decoder_output, alignments)
if self.pwgan:
input_tensor = torch.FloatTensor(postnet_output.T).unsqueeze(0)
if self.use_cuda:
input_tensor.cuda()
wav = self.pwgan.inference(
input_tensor,
hop_size=self.pwgan_ap.hop_length).data.cpu().numpy()
else:
wav = inv_spectrogram(postnet_output, self.ap, self.tts_config)
# trim silence
wav = trim_silence(wav, self.ap)
wavs += list(wav)
wavs += [0] * 10000
out = io.BytesIO()
self.save_wav(wavs, out)
return out
class MozillaTTS:
"""
Wrapper for Mozilla TTS
Related repositories:
- Mozilla TTS:
- https://github.com/mozilla/TTS
- commit 824c091
- data: https://drive.google.com/drive/folders/1FJRjGDAqWIyZRX4CsppaIPEW8UWXCWzF?usp=drive_open
- WaveRNN(optional):
- https://github.com/erogol/WaveRNN
- commit 8a1c152
- data: https://drive.google.com/drive/folders/1wpPn3a0KQc6EYtKL0qOi4NqEmhML71Ve
"""
def __init__(self, tts_model, tts_config, wavernn_model=None, wavernn_config=None, device="cpu"):
from TTS.utils.generic_utils import load_config
self.tts_config = load_config(tts_config)
self.tts_config.windowing = True
if not torch.cuda.is_available():
device = "cpu"
self.use_cuda = device != "cpu"
self.device = torch.device(device)
self.tts_model_path = tts_model
self._load_tts()
if wavernn_model and wavernn_config:
self.use_gl = False
self.batched_wavernn = True
self.wavernn_model_path = wavernn_model
self.wavernn_config = load_config(wavernn_config)
self._load_wavernn()
else:
self.use_gl = True
def _load_tts(self):
# LOAD TTS MODEL
from TTS.utils.text.symbols import symbols, phonemes
from TTS.utils.audio import AudioProcessor
from TTS.utils.generic_utils import setup_model
# load the model
num_chars = len(phonemes) if self.tts_config.use_phonemes else len(symbols)
self.tts_model = setup_model(num_chars, self.tts_config)
# load the audio processor
self._ap = AudioProcessor(**self.tts_config.audio)
# load model state
cp = torch.load(self.tts_model_path, map_location=lambda storage, loc: storage)
# load the model
self.tts_model.load_state_dict(cp['model'])
self.tts_model.to(self.device)
self.tts_model.eval()
self.tts_model.decoder.max_decoder_steps = 2000
def _load_wavernn(self):
from WaveRNN.models.wavernn import Model
self.wavernn = Model(
rnn_dims=512,
fc_dims=512,
mode="mold",
pad=2,
upsample_factors=self.wavernn_config.upsample_factors, # set this depending on dataset
feat_dims=self.wavernn_config.audio["num_mels"],
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=self._ap.hop_length,
sample_rate=self._ap.sample_rate,
).to(self.device)
check = torch.load(self.wavernn_model_path)
self.wavernn.load_state_dict(check['model'])
self.wavernn.eval()
def __call__(self, text, out_path):
waveform, alignment, mel_spec, mel_postnet_spec, stop_tokens = synthesis(self.tts_model, text, self.tts_config, self.use_cuda, self._ap, False, self.tts_config.enable_eos_bos_chars)
if not self.use_gl:
waveform = self.wavernn.generate(torch.FloatTensor(mel_postnet_spec.T).unsqueeze(0).to(self.device), batched=self.batched_wavernn, target=11000, overlap=550)
self._ap.save_wav(waveform, out_path)
upsample_factors=VOCODER_CONFIG.upsample_factors,
feat_dims=VOCODER_CONFIG.audio["num_mels"],
compute_dims=128,
res_out_dims=128,
res_blocks=10,
hop_length=ap_vocoder.hop_length,
sample_rate=ap_vocoder.sample_rate,
use_upsample_net = True,
use_aux_net = True
).cuda()
check = torch.load(VOCODER_MODEL_PATH)
wavernn.load_state_dict(check['model'], strict=False)
if use_cuda:
wavernn.cuda()
wavernn.eval();
print(check['step'])
# ### Comparision with https://mycroft.ai/blog/available-voices/
model.eval()
model.decoder.max_decoder_steps = 2000
# speaker_id = 0
for speaker_id in range(5):
sentence = "怎么网络不好啊,为什么上不去"
align, spec, stop_tokens, wav = tts(model, sentence, CONFIG, use_cuda, ap, use_gl=use_gl, figures=False, use_pinyin=use_pinyin)
sentence = "晚饭吃什么"
align, spec, stop_tokens, wav = tts(model, sentence, CONFIG, use_cuda, ap, use_gl=use_gl, figures=False, use_pinyin=use_pinyin)