def tts(
model,
vocoder_model,
text,
CONFIG,
use_cuda,
ap,
use_gl,
speaker_fileid,
speaker_embedding=None,
gst_style=None,
ap_vocoder=None,
scale_factors=None,
):
t_1 = time.time()
waveform, _, _, mel_postnet_spec, _, _ = synthesis(
model=model,
text=text,
CONFIG=CONFIG,
use_cuda=use_cuda,
ap=ap,
speaker_id=speaker_fileid,
style_wav=gst_style,
truncated=False,
enable_eos_bos_chars=CONFIG.enable_eos_bos_chars,
use_griffin_lim=use_gl,
speaker_embedding=speaker_embedding,
backend="torch",
do_trim_silence=False,
)
if CONFIG.model == "Tacotron" and not use_gl:
mel_postnet_spec = ap.out_linear_to_mel(mel_postnet_spec.T).T
mel_postnet_spec = ap._denormalize(mel_postnet_spec.T).T
if not use_gl:
vocoder_input = ap_vocoder._normalize(mel_postnet_spec.T)
if scale_factors and ap_vocoder:
# TTS and vocoder sample rates differ
_LOGGER.debug("Interpolating with scale factors %s", scale_factors)
vocoder_input = interpolate(vocoder_input, scale_factors)
else:
vocoder_input = torch.tensor(vocoder_input).unsqueeze(0)
waveform = vocoder_model.inference(vocoder_input)
if use_cuda and not use_gl:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
waveform = waveform.squeeze()
rtf = (time.time() - t_1) / (len(waveform) / ap.sample_rate)
tps = (time.time() - t_1) / len(waveform)
print(" > Run-time: {}".format(time.time() - t_1))
print(" > Real-time factor: {}".format(rtf))
print(" > Time per step: {}".format(tps))
return waveform
def tts(model, text, CONFIG, use_cuda, ap, use_gl, figures=True):
t_1 = time.time()
waveform, alignment, mel_spec, mel_postnet_spec, stop_tokens, inputs = synthesis(
model,
text,
CONFIG,
use_cuda,
ap,
speaker_id,
style_wav=None,
truncated=False,
enable_eos_bos_chars=CONFIG.enable_eos_bos_chars)
# mel_postnet_spec = ap._denormalize(mel_postnet_spec.T)
if not use_gl:
waveform = vocoder_model.inference(
torch.FloatTensor(mel_postnet_spec.T).unsqueeze(0))
waveform = waveform.flatten()
if use_cuda:
waveform = waveform.cpu()
waveform = waveform.numpy()
rtf = (time.time() - t_1) / (len(waveform) / ap.sample_rate)
tps = (time.time() - t_1) / len(waveform)
print(waveform.shape)
print(" > Run-time: {}".format(time.time() - t_1))
print(" > Real-time factor: {}".format(rtf))
print(" > Time per step: {}".format(tps))
IPython.display.display(
IPython.display.Audio(waveform, rate=CONFIG.audio['sample_rate']))
return alignment, mel_postnet_spec, stop_tokens, waveform
def test_run(self, ap) -> Tuple[Dict, Dict]:
"""Generic test run for `tts` models used by `Trainer`.
You can override this for a different behaviour.
Returns:
Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard.
"""
print(" | > Synthesizing test sentences.")
test_audios = {}
test_figures = {}
test_sentences = self.config.test_sentences
aux_inputs = self.get_aux_input()
for idx, sen in enumerate(test_sentences):
outputs_dict = synthesis(
self,
sen,
self.config,
"cuda" in str(next(self.parameters()).device),
ap,
speaker_id=aux_inputs["speaker_id"],
d_vector=aux_inputs["d_vector"],
style_wav=aux_inputs["style_wav"],
enable_eos_bos_chars=self.config.enable_eos_bos_chars,
use_griffin_lim=True,
do_trim_silence=False,
)
test_audios["{}-audio".format(idx)] = outputs_dict["wav"]
test_figures["{}-prediction".format(idx)] = plot_spectrogram(
outputs_dict["outputs"]["model_outputs"], ap, output_fig=False
)
test_figures["{}-alignment".format(idx)] = plot_alignment(
outputs_dict["outputs"]["alignments"], output_fig=False
)
return test_figures, test_audios
def tts(self, text, speaker_idx=None):
start_time = time.time()
wavs = []
sens = self.split_into_sentences(text)
print(" > Text splitted to sentences.")
print(sens)
speaker_embedding = self.init_speaker(speaker_idx)
use_gl = self.vocoder_model is None
for sen in sens:
# synthesize voice
waveform, _, _, mel_postnet_spec, _, _ = synthesis(
self.tts_model,
sen,
self.tts_config,
self.use_cuda,
self.ap,
speaker_idx,
None,
False,
self.tts_config.enable_eos_bos_chars,
use_gl,
speaker_embedding=speaker_embedding)
if not use_gl:
# denormalize tts output based on tts audio config
mel_postnet_spec = self.ap.denormalize(mel_postnet_spec.T).T
device_type = "cuda" if self.use_cuda else "cpu"
# renormalize spectrogram based on vocoder config
vocoder_input = self.vocoder_ap.normalize(mel_postnet_spec.T)
# compute scale factor for possible sample rate mismatch
scale_factor = [1, self.vocoder_config['audio']['sample_rate'] / self.ap.sample_rate]
if scale_factor[1] != 1:
print(" > interpolating tts model output.")
vocoder_input = interpolate_vocoder_input(scale_factor, vocoder_input)
else:
vocoder_input = torch.tensor(vocoder_input).unsqueeze(0) # pylint: disable=not-callable
# run vocoder model
# [1, T, C]
waveform = self.vocoder_model.inference(vocoder_input.to(device_type))
if self.use_cuda and not use_gl:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
waveform = waveform.squeeze()
# trim silence
waveform = trim_silence(waveform, self.ap)
wavs += list(waveform)
wavs += [0] * 10000
# 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 wavs
def text_to_speech(text):
t_1 = time.time()
waveform, alignment, mel_spec, mel_postnet_spec, stop_tokens, inputs = synthesis(model, text, TTS_CONFIG, False, ap, speaker_id, style_wav=None, truncated=False, enable_eos_bos_chars=TTS_CONFIG.enable_eos_bos_chars)
# mel_postnet_spec = ap._denormalize(mel_postnet_spec.T)
waveform = vocoder_model.inference(torch.FloatTensor(mel_postnet_spec.T).unsqueeze(0))
waveform = waveform.flatten()
waveform = waveform.numpy()
waveform = (waveform * 32768).astype(np.int16, order='C')
buf = io.BytesIO()
scipy.io.wavfile.write(buf, TTS_CONFIG.audio['sample_rate'], waveform)
audio = buf.read()
print("Converted text \"{}\" to {:.2f}s / {}kB of audio".format(text[:40] + (text[40:] and "..."), len(waveform) / ap.sample_rate, len(audio) // 1024))
return audio
def tts(model, vocoder_model, text, CONFIG, use_cuda, ap, use_gl, speaker_fileid, speaker_embedding=None, gst_style=None):
t_1 = time.time()
waveform, _, _, mel_postnet_spec, _, _ = synthesis(model, text, CONFIG, use_cuda, ap, speaker_fileid, gst_style, False, CONFIG.enable_eos_bos_chars, use_gl, speaker_embedding=speaker_embedding)
if CONFIG.model == "Tacotron" and not use_gl:
mel_postnet_spec = ap.out_linear_to_mel(mel_postnet_spec.T).T
if not use_gl:
waveform = vocoder_model.inference(torch.FloatTensor(mel_postnet_spec.T).unsqueeze(0))
if use_cuda and not use_gl:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
waveform = waveform.squeeze()
rtf = (time.time() - t_1) / (len(waveform) / ap.sample_rate)
tps = (time.time() - t_1) / len(waveform)
print(" > Run-time: {}".format(time.time() - t_1))
print(" > Real-time factor: {}".format(rtf))
print(" > Time per step: {}".format(tps))
return waveform
def text_to_speech(self,
text,
length_scale=1.1,
noise_scale=0.4,
speaker_id=None,
use_gl=False):
# run tts
target_sr = TTS_CONFIG.audio['sample_rate']
waveform, alignment, mel_spec, mel_postnet_spec, stop_tokens, inputs =\
synthesis(self.model,
text,
TTS_CONFIG,
USE_CUDA,
self.audio_processor,
speaker_id,
None,
False,
TTS_CONFIG.enable_eos_bos_chars,
use_gl)
# run vocoder
mel_postnet_spec = self.audio_processor._denormalize(
mel_postnet_spec.T).T
if not use_gl:
target_sr = VOCODER_CONFIG.audio['sample_rate']
vocoder_input = self.ap_vocoder._normalize(mel_postnet_spec.T)
if self.scale_factor[1] != 1:
vocoder_input = interpolate_vocoder_input(
self.scale_factor, vocoder_input)
else:
vocoder_input = torch.tensor(vocoder_input).unsqueeze(0)
waveform = self.vocoder_model.inference(vocoder_input)
# format output
if USE_CUDA and not use_gl:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
waveform = waveform.squeeze()
return waveform
def test_run(self, assets: Dict) -> Tuple[Dict, Dict]:
"""Generic test run for `tts` models used by `Trainer`.
You can override this for a different behaviour.
Args:
assets (dict): A dict of training assets. For `tts` models, it must include `{'audio_processor': ap}`.
Returns:
Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard.
"""
print(" | > Synthesizing test sentences.")
test_audios = {}
test_figures = {}
test_sentences = self.config.test_sentences
aux_inputs = self._get_test_aux_input()
for idx, sen in enumerate(test_sentences):
outputs_dict = synthesis(
self,
sen,
self.config,
"cuda" in str(next(self.parameters()).device),
speaker_id=aux_inputs["speaker_id"],
d_vector=aux_inputs["d_vector"],
style_wav=aux_inputs["style_wav"],
use_griffin_lim=True,
do_trim_silence=False,
)
test_audios["{}-audio".format(idx)] = outputs_dict["wav"]
test_figures["{}-prediction".format(idx)] = plot_spectrogram(
outputs_dict["outputs"]["model_outputs"], self.ap, output_fig=False
)
test_figures["{}-alignment".format(idx)] = plot_alignment(
outputs_dict["outputs"]["alignments"], output_fig=False
)
return {"figures": test_figures, "audios": test_audios}
def test_run(self, ap) -> Tuple[Dict, Dict]:
"""Generic test run for `tts` models used by `Trainer`.
You can override this for a different behaviour.
Returns:
Tuple[Dict, Dict]: Test figures and audios to be projected to Tensorboard.
"""
print(" | > Synthesizing test sentences.")
test_audios = {}
test_figures = {}
test_sentences = self.config.test_sentences
for idx, s_info in enumerate(test_sentences):
try:
aux_inputs = self.get_aux_input_from_test_sentences(s_info)
wav, alignment, _, _ = synthesis(
self,
aux_inputs["text"],
self.config,
"cuda" in str(next(self.parameters()).device),
ap,
speaker_id=aux_inputs["speaker_id"],
d_vector=aux_inputs["d_vector"],
style_wav=aux_inputs["style_wav"],
language_id=aux_inputs["language_id"],
language_name=aux_inputs["language_name"],
enable_eos_bos_chars=self.config.enable_eos_bos_chars,
use_griffin_lim=True,
do_trim_silence=False,
).values()
test_audios["{}-audio".format(idx)] = wav
test_figures["{}-alignment".format(idx)] = plot_alignment(
alignment.T, output_fig=False)
except: # pylint: disable=bare-except
print(" !! Error creating Test Sentence -", idx)
return test_figures, test_audios
def evaluate(data_loader, model, criterion, ap, global_step, epoch):
model.eval()
epoch_time = 0
keep_avg = KeepAverage()
c_logger.print_eval_start()
if data_loader is not None:
for num_iter, data in enumerate(data_loader):
start_time = time.time()
# format data
text_input, text_lengths, mel_targets, mel_lengths, speaker_c,\
_, _, _, dur_target, _ = format_data(data)
# forward pass model
with torch.cuda.amp.autocast(enabled=c.mixed_precision):
decoder_output, dur_output, alignments = model.forward(
text_input,
text_lengths,
mel_lengths,
dur_target,
g=speaker_c)
# compute loss
loss_dict = criterion(decoder_output, mel_targets,
mel_lengths, dur_output,
torch.log(1 + dur_target), text_lengths)
# step time
step_time = time.time() - start_time
epoch_time += step_time
# compute alignment score
align_error = 1 - alignment_diagonal_score(alignments, binary=True)
loss_dict['align_error'] = align_error
# aggregate losses from processes
if num_gpus > 1:
loss_dict['loss_l1'] = reduce_tensor(loss_dict['loss_l1'].data,
num_gpus)
loss_dict['loss_ssim'] = reduce_tensor(
loss_dict['loss_ssim'].data, num_gpus)
loss_dict['loss_dur'] = reduce_tensor(
loss_dict['loss_dur'].data, num_gpus)
loss_dict['loss'] = reduce_tensor(loss_dict['loss'].data,
num_gpus)
# detach loss values
loss_dict_new = dict()
for key, value in loss_dict.items():
if isinstance(value, (int, float)):
loss_dict_new[key] = value
else:
loss_dict_new[key] = value.item()
loss_dict = loss_dict_new
# update avg stats
update_train_values = dict()
for key, value in loss_dict.items():
update_train_values['avg_' + key] = value
keep_avg.update_values(update_train_values)
if c.print_eval:
c_logger.print_eval_step(num_iter, loss_dict,
keep_avg.avg_values)
if args.rank == 0:
# Diagnostic visualizations
idx = np.random.randint(mel_targets.shape[0])
pred_spec = decoder_output[idx].detach().data.cpu().numpy().T
gt_spec = mel_targets[idx].data.cpu().numpy().T
align_img = alignments[idx].data.cpu()
eval_figures = {
"prediction": plot_spectrogram(pred_spec, ap,
output_fig=False),
"ground_truth": plot_spectrogram(gt_spec, ap,
output_fig=False),
"alignment": plot_alignment(align_img, output_fig=False)
}
# Sample audio
eval_audio = ap.inv_melspectrogram(pred_spec.T)
tb_logger.tb_eval_audios(global_step, {"ValAudio": eval_audio},
c.audio["sample_rate"])
# Plot Validation Stats
tb_logger.tb_eval_stats(global_step, keep_avg.avg_values)
tb_logger.tb_eval_figures(global_step, eval_figures)
if args.rank == 0 and epoch >= c.test_delay_epochs and epoch % c.test_every_epochs == 0:
if c.test_sentences_file is None:
test_sentences = [
"ජනක ප්රදීප් ලියනගේ.",
"රගර් ගැහුවා කියල කොහොමද බූරුවො වොලි බෝල් නැති වෙන්නෙ.",
"රට්ඨපාල කුමරු ගිහිගෙය හැර පැවිදි වී සිටියි.",
"අජාසත් රජතුමාගේ ඇත් සේනාවේ අති භයානක ඇතෙක් සිටියා."
]
else:
with open(c.test_sentences_file, "r") as f:
test_sentences = [s.strip() for s in f.readlines()]
# test sentences
test_audios = {}
test_figures = {}
print(" | > Synthesizing test sentences")
if c.use_speaker_embedding:
if c.use_external_speaker_embedding_file:
speaker_embedding = speaker_mapping[list(
speaker_mapping.keys())[randrange(
len(speaker_mapping) - 1)]]['embedding']
speaker_id = None
else:
speaker_id = 0
speaker_embedding = None
else:
speaker_id = None
speaker_embedding = None
style_wav = c.get("style_wav_for_test")
for idx, test_sentence in enumerate(test_sentences):
try:
wav, alignment, _, postnet_output, _, _ = synthesis(
model,
test_sentence,
c,
use_cuda,
ap,
speaker_id=speaker_id,
speaker_embedding=speaker_embedding,
style_wav=style_wav,
truncated=False,
enable_eos_bos_chars=c.enable_eos_bos_chars, #pylint: disable=unused-argument
use_griffin_lim=True,
do_trim_silence=False)
file_path = os.path.join(AUDIO_PATH, str(global_step))
os.makedirs(file_path, exist_ok=True)
file_path = os.path.join(file_path,
"TestSentence_{}.wav".format(idx))
ap.save_wav(wav, file_path)
test_audios['{}-audio'.format(idx)] = wav
test_figures['{}-prediction'.format(idx)] = plot_spectrogram(
postnet_output, ap)
test_figures['{}-alignment'.format(idx)] = plot_alignment(
alignment)
except: #pylint: disable=bare-except
print(" !! Error creating Test Sentence -", idx)
traceback.print_exc()
tb_logger.tb_test_audios(global_step, test_audios,
c.audio['sample_rate'])
tb_logger.tb_test_figures(global_step, test_figures)
return keep_avg.avg_values
def tts(
self,
text: str = "",
speaker_name: str = "",
language_name: str = "",
speaker_wav: Union[str, List[str]] = None,
style_wav=None,
reference_wav=None,
reference_speaker_name=None,
) -> List[int]:
"""🐸 TTS magic. Run all the models and generate speech.
Args:
text (str): input text.
speaker_name (str, optional): spekaer id for multi-speaker models. Defaults to "".
language_name (str, optional): language id for multi-language models. Defaults to "".
speaker_wav (Union[str, List[str]], optional): path to the speaker wav. Defaults to None.
style_wav ([type], optional): style waveform for GST. Defaults to None.
reference_wav ([type], optional): reference waveform for voice conversion. Defaults to None.
reference_speaker_name ([type], optional): spekaer id of reference waveform. Defaults to None.
Returns:
List[int]: [description]
"""
start_time = time.time()
wavs = []
if not text and not reference_wav:
raise ValueError(
"You need to define either `text` (for sythesis) or a `reference_wav` (for voice conversion) to use the Coqui TTS API."
)
if text:
sens = self.split_into_sentences(text)
print(" > Text splitted to sentences.")
print(sens)
# handle multi-speaker
speaker_embedding = None
speaker_id = None
if self.tts_speakers_file or hasattr(self.tts_model.speaker_manager,
"ids"):
if speaker_name and isinstance(speaker_name, str):
if self.tts_config.use_d_vector_file:
# get the average speaker embedding from the saved d_vectors.
speaker_embedding = self.tts_model.speaker_manager.get_mean_embedding(
speaker_name, num_samples=None, randomize=False)
speaker_embedding = np.array(speaker_embedding)[
None, :] # [1 x embedding_dim]
else:
# get speaker idx from the speaker name
speaker_id = self.tts_model.speaker_manager.ids[
speaker_name]
elif not speaker_name and not speaker_wav:
raise ValueError(
" [!] Look like you use a multi-speaker model. "
"You need to define either a `speaker_name` or a `speaker_wav` to use a multi-speaker model."
)
else:
speaker_embedding = None
else:
if speaker_name:
raise ValueError(
f" [!] Missing speakers.json file path for selecting speaker {speaker_name}."
"Define path for speaker.json if it is a multi-speaker model or remove defined speaker idx. "
)
# handle multi-lingaul
language_id = None
if self.tts_languages_file or (
hasattr(self.tts_model, "language_manager")
and self.tts_model.language_manager is not None):
if language_name and isinstance(language_name, str):
language_id = self.tts_model.language_manager.ids[
language_name]
elif not language_name:
raise ValueError(
" [!] Look like you use a multi-lingual model. "
"You need to define either a `language_name` or a `style_wav` to use a multi-lingual model."
)
else:
raise ValueError(
f" [!] Missing language_ids.json file path for selecting language {language_name}."
"Define path for language_ids.json if it is a multi-lingual model or remove defined language idx. "
)
# compute a new d_vector from the given clip.
if speaker_wav is not None:
speaker_embedding = self.tts_model.speaker_manager.compute_embedding_from_clip(
speaker_wav)
use_gl = self.vocoder_model is None
if not reference_wav:
for sen in sens:
# synthesize voice
outputs = synthesis(
model=self.tts_model,
text=sen,
CONFIG=self.tts_config,
use_cuda=self.use_cuda,
speaker_id=speaker_id,
language_id=language_id,
style_wav=style_wav,
use_griffin_lim=use_gl,
d_vector=speaker_embedding,
)
waveform = outputs["wav"]
mel_postnet_spec = outputs["outputs"]["model_outputs"][
0].detach().cpu().numpy()
if not use_gl:
# denormalize tts output based on tts audio config
mel_postnet_spec = self.tts_model.ap.denormalize(
mel_postnet_spec.T).T
device_type = "cuda" if self.use_cuda else "cpu"
# renormalize spectrogram based on vocoder config
vocoder_input = self.vocoder_ap.normalize(
mel_postnet_spec.T)
# compute scale factor for possible sample rate mismatch
scale_factor = [
1,
self.vocoder_config["audio"]["sample_rate"] /
self.tts_model.ap.sample_rate,
]
if scale_factor[1] != 1:
print(" > interpolating tts model output.")
vocoder_input = interpolate_vocoder_input(
scale_factor, vocoder_input)
else:
vocoder_input = torch.tensor(vocoder_input).unsqueeze(
0) # pylint: disable=not-callable
# run vocoder model
# [1, T, C]
waveform = self.vocoder_model.inference(
vocoder_input.to(device_type))
if self.use_cuda and not use_gl:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
waveform = waveform.squeeze()
# trim silence
if self.tts_config.audio["do_trim_silence"] is True:
waveform = trim_silence(waveform, self.tts_model.ap)
wavs += list(waveform)
wavs += [0] * 10000
else:
# get the speaker embedding or speaker id for the reference wav file
reference_speaker_embedding = None
reference_speaker_id = None
if self.tts_speakers_file or hasattr(
self.tts_model.speaker_manager, "speaker_ids"):
if reference_speaker_name and isinstance(
reference_speaker_name, str):
if self.tts_config.use_d_vector_file:
# get the speaker embedding from the saved d_vectors.
reference_speaker_embedding = self.tts_model.speaker_manager.get_embeddings_by_name(
reference_speaker_name)[0]
reference_speaker_embedding = np.array(
reference_speaker_embedding)[
None, :] # [1 x embedding_dim]
else:
# get speaker idx from the speaker name
reference_speaker_id = self.tts_model.speaker_manager.ids[
reference_speaker_name]
else:
reference_speaker_embedding = self.tts_model.speaker_manager.compute_embedding_from_clip(
reference_wav)
outputs = transfer_voice(
model=self.tts_model,
CONFIG=self.tts_config,
use_cuda=self.use_cuda,
reference_wav=reference_wav,
speaker_id=speaker_id,
d_vector=speaker_embedding,
use_griffin_lim=use_gl,
reference_speaker_id=reference_speaker_id,
reference_d_vector=reference_speaker_embedding,
)
waveform = outputs
if not use_gl:
mel_postnet_spec = outputs[0].detach().cpu().numpy()
# denormalize tts output based on tts audio config
mel_postnet_spec = self.tts_model.ap.denormalize(
mel_postnet_spec.T).T
device_type = "cuda" if self.use_cuda else "cpu"
# renormalize spectrogram based on vocoder config
vocoder_input = self.vocoder_ap.normalize(mel_postnet_spec.T)
# compute scale factor for possible sample rate mismatch
scale_factor = [
1,
self.vocoder_config["audio"]["sample_rate"] /
self.tts_model.ap.sample_rate,
]
if scale_factor[1] != 1:
print(" > interpolating tts model output.")
vocoder_input = interpolate_vocoder_input(
scale_factor, vocoder_input)
else:
vocoder_input = torch.tensor(vocoder_input).unsqueeze(0) # pylint: disable=not-callable
# run vocoder model
# [1, T, C]
waveform = self.vocoder_model.inference(
vocoder_input.to(device_type))
if self.use_cuda:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
wavs = waveform.squeeze()
# 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 wavs
def evaluate(data_loader, model, criterion, ap, global_step, epoch):
model.eval()
epoch_time = 0
keep_avg = KeepAverage()
c_logger.print_eval_start()
if data_loader is not None:
for num_iter, data in enumerate(data_loader):
start_time = time.time()
# format data
text_input, text_lengths, mel_targets, mel_lengths, speaker_c, _, _, _, dur_target, _ = format_data(data)
# forward pass model
with torch.cuda.amp.autocast(enabled=c.mixed_precision):
decoder_output, dur_output, alignments = model.forward(
text_input, text_lengths, mel_lengths, dur_target, g=speaker_c
)
# compute loss
loss_dict = criterion(
decoder_output, mel_targets, mel_lengths, dur_output, torch.log(1 + dur_target), text_lengths
)
# step time
step_time = time.time() - start_time
epoch_time += step_time
# compute alignment score
align_error = 1 - alignment_diagonal_score(alignments, binary=True)
loss_dict["align_error"] = align_error
# aggregate losses from processes
if num_gpus > 1:
loss_dict["loss_l1"] = reduce_tensor(loss_dict["loss_l1"].data, num_gpus)
loss_dict["loss_ssim"] = reduce_tensor(loss_dict["loss_ssim"].data, num_gpus)
loss_dict["loss_dur"] = reduce_tensor(loss_dict["loss_dur"].data, num_gpus)
loss_dict["loss"] = reduce_tensor(loss_dict["loss"].data, num_gpus)
# detach loss values
loss_dict_new = dict()
for key, value in loss_dict.items():
if isinstance(value, (int, float)):
loss_dict_new[key] = value
else:
loss_dict_new[key] = value.item()
loss_dict = loss_dict_new
# update avg stats
update_train_values = dict()
for key, value in loss_dict.items():
update_train_values["avg_" + key] = value
keep_avg.update_values(update_train_values)
if c.print_eval:
c_logger.print_eval_step(num_iter, loss_dict, keep_avg.avg_values)
if args.rank == 0:
# Diagnostic visualizations
idx = np.random.randint(mel_targets.shape[0])
pred_spec = decoder_output[idx].detach().data.cpu().numpy().T
gt_spec = mel_targets[idx].data.cpu().numpy().T
align_img = alignments[idx].data.cpu()
eval_figures = {
"prediction": plot_spectrogram(pred_spec, ap, output_fig=False),
"ground_truth": plot_spectrogram(gt_spec, ap, output_fig=False),
"alignment": plot_alignment(align_img, output_fig=False),
}
# Sample audio
eval_audio = ap.inv_melspectrogram(pred_spec.T)
tb_logger.tb_eval_audios(global_step, {"ValAudio": eval_audio}, c.audio["sample_rate"])
# Plot Validation Stats
tb_logger.tb_eval_stats(global_step, keep_avg.avg_values)
tb_logger.tb_eval_figures(global_step, eval_figures)
if args.rank == 0 and epoch >= c.test_delay_epochs:
if c.test_sentences_file is None:
test_sentences = [
"It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
"Be a voice, not an echo.",
"I'm sorry Dave. I'm afraid I can't do that.",
"This cake is great. It's so delicious and moist.",
"Prior to November 22, 1963.",
]
else:
with open(c.test_sentences_file, "r") as f:
test_sentences = [s.strip() for s in f.readlines()]
# test sentences
test_audios = {}
test_figures = {}
print(" | > Synthesizing test sentences")
if c.use_speaker_embedding:
if c.use_external_speaker_embedding_file:
speaker_embedding = speaker_mapping[list(speaker_mapping.keys())[randrange(len(speaker_mapping) - 1)]][
"embedding"
]
speaker_id = None
else:
speaker_id = 0
speaker_embedding = None
else:
speaker_id = None
speaker_embedding = None
style_wav = c.get("style_wav_for_test")
for idx, test_sentence in enumerate(test_sentences):
try:
wav, alignment, _, postnet_output, _, _ = synthesis(
model,
test_sentence,
c,
use_cuda,
ap,
speaker_id=speaker_id,
speaker_embedding=speaker_embedding,
style_wav=style_wav,
truncated=False,
enable_eos_bos_chars=c.enable_eos_bos_chars, # pylint: disable=unused-argument
use_griffin_lim=True,
do_trim_silence=False,
)
file_path = os.path.join(AUDIO_PATH, str(global_step))
os.makedirs(file_path, exist_ok=True)
file_path = os.path.join(file_path, "TestSentence_{}.wav".format(idx))
ap.save_wav(wav, file_path)
test_audios["{}-audio".format(idx)] = wav
test_figures["{}-prediction".format(idx)] = plot_spectrogram(postnet_output, ap)
test_figures["{}-alignment".format(idx)] = plot_alignment(alignment)
except: # pylint: disable=bare-except
print(" !! Error creating Test Sentence -", idx)
traceback.print_exc()
tb_logger.tb_test_audios(global_step, test_audios, c.audio["sample_rate"])
tb_logger.tb_test_figures(global_step, test_figures)
return keep_avg.avg_values
def evaluate(data_loader, model, criterion, ap, global_step, epoch):
model.eval()
epoch_time = 0
keep_avg = KeepAverage()
c_logger.print_eval_start()
if data_loader is not None:
for num_iter, data in enumerate(data_loader):
start_time = time.time()
# format data
(
text_input,
text_lengths,
mel_input,
mel_lengths,
linear_input,
stop_targets,
speaker_ids,
speaker_embeddings,
_,
_,
) = format_data(data)
assert mel_input.shape[1] % model.decoder.r == 0
# forward pass model
if c.bidirectional_decoder or c.double_decoder_consistency:
(
decoder_output,
postnet_output,
alignments,
stop_tokens,
decoder_backward_output,
alignments_backward,
) = model(text_input,
text_lengths,
mel_input,
speaker_ids=speaker_ids,
speaker_embeddings=speaker_embeddings)
else:
decoder_output, postnet_output, alignments, stop_tokens = model(
text_input,
text_lengths,
mel_input,
speaker_ids=speaker_ids,
speaker_embeddings=speaker_embeddings)
decoder_backward_output = None
alignments_backward = None
# set the alignment lengths wrt reduction factor for guided attention
if mel_lengths.max() % model.decoder.r != 0:
alignment_lengths = (
mel_lengths +
(model.decoder.r -
(mel_lengths.max() % model.decoder.r))) // model.decoder.r
else:
alignment_lengths = mel_lengths // model.decoder.r
# compute loss
loss_dict = criterion(
postnet_output,
decoder_output,
mel_input,
linear_input,
stop_tokens,
stop_targets,
mel_lengths,
decoder_backward_output,
alignments,
alignment_lengths,
alignments_backward,
text_lengths,
)
# step time
step_time = time.time() - start_time
epoch_time += step_time
# compute alignment score
align_error = 1 - alignment_diagonal_score(alignments)
loss_dict["align_error"] = align_error
# aggregate losses from processes
if num_gpus > 1:
loss_dict["postnet_loss"] = reduce_tensor(
loss_dict["postnet_loss"].data, num_gpus)
loss_dict["decoder_loss"] = reduce_tensor(
loss_dict["decoder_loss"].data, num_gpus)
if c.stopnet:
loss_dict["stopnet_loss"] = reduce_tensor(
loss_dict["stopnet_loss"].data, num_gpus)
# detach loss values
loss_dict_new = dict()
for key, value in loss_dict.items():
if isinstance(value, (int, float)):
loss_dict_new[key] = value
else:
loss_dict_new[key] = value.item()
loss_dict = loss_dict_new
# update avg stats
update_train_values = dict()
for key, value in loss_dict.items():
update_train_values["avg_" + key] = value
keep_avg.update_values(update_train_values)
if c.print_eval:
c_logger.print_eval_step(num_iter, loss_dict,
keep_avg.avg_values)
if args.rank == 0:
# Diagnostic visualizations
idx = np.random.randint(mel_input.shape[0])
const_spec = postnet_output[idx].data.cpu().numpy()
gt_spec = (linear_input[idx].data.cpu().numpy() if c.model in [
"Tacotron", "TacotronGST"
] else mel_input[idx].data.cpu().numpy())
align_img = alignments[idx].data.cpu().numpy()
eval_figures = {
"prediction": plot_spectrogram(const_spec,
ap,
output_fig=False),
"ground_truth": plot_spectrogram(gt_spec, ap,
output_fig=False),
"alignment": plot_alignment(align_img, output_fig=False),
}
# Sample audio
if c.model in ["Tacotron", "TacotronGST"]:
eval_audio = ap.inv_spectrogram(const_spec.T)
else:
eval_audio = ap.inv_melspectrogram(const_spec.T)
tb_logger.tb_eval_audios(global_step, {"ValAudio": eval_audio},
c.audio["sample_rate"])
# Plot Validation Stats
if c.bidirectional_decoder or c.double_decoder_consistency:
align_b_img = alignments_backward[idx].data.cpu().numpy()
eval_figures["alignment2"] = plot_alignment(align_b_img,
output_fig=False)
tb_logger.tb_eval_stats(global_step, keep_avg.avg_values)
tb_logger.tb_eval_figures(global_step, eval_figures)
if args.rank == 0 and epoch > c.test_delay_epochs:
if c.test_sentences_file is None:
test_sentences = [
"It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
"Be a voice, not an echo.",
"I'm sorry Dave. I'm afraid I can't do that.",
"This cake is great. It's so delicious and moist.",
"Prior to November 22, 1963.",
]
else:
with open(c.test_sentences_file, "r") as f:
test_sentences = [s.strip() for s in f.readlines()]
# test sentences
test_audios = {}
test_figures = {}
print(" | > Synthesizing test sentences")
speaker_id = 0 if c.use_speaker_embedding else None
speaker_embedding = (speaker_mapping[list(
speaker_mapping.keys())[randrange(len(speaker_mapping) -
1)]]["embedding"]
if c.use_external_speaker_embedding_file
and c.use_speaker_embedding else None)
style_wav = c.get("gst_style_input")
if style_wav is None and c.use_gst:
# inicialize GST with zero dict.
style_wav = {}
print(
"WARNING: You don't provided a gst style wav, for this reason we use a zero tensor!"
)
for i in range(c.gst["gst_style_tokens"]):
style_wav[str(i)] = 0
style_wav = c.get("gst_style_input")
for idx, test_sentence in enumerate(test_sentences):
try:
wav, alignment, decoder_output, postnet_output, stop_tokens, _ = synthesis(
model,
test_sentence,
c,
use_cuda,
ap,
speaker_id=speaker_id,
speaker_embedding=speaker_embedding,
style_wav=style_wav,
truncated=False,
enable_eos_bos_chars=c.enable_eos_bos_chars, # pylint: disable=unused-argument
use_griffin_lim=True,
do_trim_silence=False,
)
file_path = os.path.join(AUDIO_PATH, str(global_step))
os.makedirs(file_path, exist_ok=True)
file_path = os.path.join(file_path,
"TestSentence_{}.wav".format(idx))
ap.save_wav(wav, file_path)
test_audios["{}-audio".format(idx)] = wav
test_figures["{}-prediction".format(idx)] = plot_spectrogram(
postnet_output, ap, output_fig=False)
test_figures["{}-alignment".format(idx)] = plot_alignment(
alignment, output_fig=False)
except: # pylint: disable=bare-except
print(" !! Error creating Test Sentence -", idx)
traceback.print_exc()
tb_logger.tb_test_audios(global_step, test_audios,
c.audio["sample_rate"])
tb_logger.tb_test_figures(global_step, test_figures)
return keep_avg.avg_values
def tts(self,
text: str,
speaker_idx: str = "",
speaker_wav=None,
style_wav=None) -> List[int]:
"""🐸 TTS magic. Run all the models and generate speech.
Args:
text (str): input text.
speaker_idx (str, optional): spekaer id for multi-speaker models. Defaults to "".
speaker_wav ():
style_wav ([type], optional): style waveform for GST. Defaults to None.
Returns:
List[int]: [description]
"""
start_time = time.time()
wavs = []
sens = self.split_into_sentences(text)
print(" > Text splitted to sentences.")
print(sens)
# handle multi-speaker
speaker_embedding = None
speaker_id = None
if isinstance(speaker_idx, int):
speaker_id = speaker_idx
elif self.tts_speakers_file:
if speaker_idx and isinstance(speaker_idx, str):
if self.tts_config.use_d_vector_file:
# get the speaker embedding from the saved d_vectors.
speaker_embedding = self.tts_model.speaker_manager.get_d_vectors_by_speaker(
speaker_idx)[0]
else:
# get speaker idx from the speaker name
try:
speaker_id = self.tts_model.speaker_manager.speaker_ids[
speaker_idx]
except KeyError:
# Interpet as int
speaker_id = int(speaker_idx)
elif not speaker_idx and not speaker_wav:
raise ValueError(
" [!] Look like you use a multi-speaker model. "
"You need to define either a `speaker_idx` or a `style_wav` to use a multi-speaker model."
)
else:
speaker_embedding = None
else:
if speaker_idx:
raise ValueError(
f" [!] Missing speaker.json file path for selecting speaker {speaker_idx}."
"Define path for speaker.json if it is a multi-speaker model or remove defined speaker idx. "
)
# compute a new d_vector from the given clip.
if speaker_wav is not None:
speaker_embedding = self.tts_model.speaker_manager.compute_d_vector_from_clip(
speaker_wav)
use_gl = self.vocoder_model is None
for sen in sens:
# synthesize voice
outputs = synthesis(
model=self.tts_model,
text=sen,
CONFIG=self.tts_config,
use_cuda=self.use_cuda,
ap=self.ap,
speaker_id=speaker_id,
style_wav=style_wav,
enable_eos_bos_chars=self.tts_config.enable_eos_bos_chars,
use_griffin_lim=use_gl,
d_vector=speaker_embedding,
)
waveform = outputs["wav"]
mel_postnet_spec = (
outputs["outputs"]["model_outputs"][0].detach().cpu().numpy())
if not use_gl:
# denormalize tts output based on tts audio config
mel_postnet_spec = self.ap.denormalize(mel_postnet_spec.T).T
device_type = "cuda" if self.use_cuda else "cpu"
# renormalize spectrogram based on vocoder config
vocoder_input = self.vocoder_ap.normalize(mel_postnet_spec.T)
# compute scale factor for possible sample rate mismatch
scale_factor = [
1,
self.vocoder_config["audio"]["sample_rate"] /
self.ap.sample_rate,
]
if scale_factor[1] != 1:
print(" > interpolating tts model output.")
vocoder_input = interpolate_vocoder_input(
scale_factor, vocoder_input)
else:
vocoder_input = torch.tensor(vocoder_input).unsqueeze(0) # pylint: disable=not-callable
# run vocoder model
# [1, T, C]
waveform = self.vocoder_model.inference(
vocoder_input.to(device_type))
if self.use_cuda and not use_gl:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
waveform = waveform.squeeze()
# trim silence
waveform = trim_silence(waveform, self.ap)
wavs += list(waveform)
wavs += [0] * 10000
# 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 wavs
def evaluate(model, criterion, ap, global_step, epoch, speaker_mapping=None):
data_loader = setup_loader(ap,
model.decoder.r,
is_val=True,
speaker_mapping=speaker_mapping)
model.eval()
epoch_time = 0
keep_avg = KeepAverage()
c_logger.print_eval_start()
if data_loader is not None:
for num_iter, data in enumerate(data_loader):
start_time = time.time()
# format data
text_input, text_lengths, mel_input, mel_lengths, linear_input, stop_targets, speaker_ids, speaker_embeddings, _, _ = format_data(
data, speaker_mapping)
assert mel_input.shape[1] % model.decoder.r == 0
# forward pass model
if c.bidirectional_decoder or c.double_decoder_consistency:
decoder_output, postnet_output, alignments, stop_tokens, decoder_backward_output, alignments_backward = model(
text_input,
text_lengths,
mel_input,
speaker_ids=speaker_ids,
speaker_embeddings=speaker_embeddings)
else:
decoder_output, postnet_output, alignments, stop_tokens = model(
text_input,
text_lengths,
mel_input,
speaker_ids=speaker_ids,
speaker_embeddings=speaker_embeddings)
decoder_backward_output = None
alignments_backward = None
# set the alignment lengths wrt reduction factor for guided attention
if mel_lengths.max() % model.decoder.r != 0:
alignment_lengths = (
mel_lengths +
(model.decoder.r -
(mel_lengths.max() % model.decoder.r))) // model.decoder.r
else:
alignment_lengths = mel_lengths // model.decoder.r
# compute loss
loss_dict = criterion(postnet_output, decoder_output, mel_input,
linear_input, stop_tokens, stop_targets,
mel_lengths, decoder_backward_output,
alignments, alignment_lengths,
alignments_backward, text_lengths)
# step time
step_time = time.time() - start_time
epoch_time += step_time
# compute alignment score
align_error = 1 - alignment_diagonal_score(alignments)
loss_dict['align_error'] = align_error
# aggregate losses from processes
if num_gpus > 1:
loss_dict['postnet_loss'] = reduce_tensor(
loss_dict['postnet_loss'].data, num_gpus)
loss_dict['decoder_loss'] = reduce_tensor(
loss_dict['decoder_loss'].data, num_gpus)
if c.stopnet:
loss_dict['stopnet_loss'] = reduce_tensor(
loss_dict['stopnet_loss'].data, num_gpus)
# detach loss values
loss_dict_new = dict()
for key, value in loss_dict.items():
if isinstance(value, (int, float)):
loss_dict_new[key] = value
else:
loss_dict_new[key] = value.item()
loss_dict = loss_dict_new
# update avg stats
update_train_values = dict()
for key, value in loss_dict.items():
update_train_values['avg_' + key] = value
keep_avg.update_values(update_train_values)
if c.print_eval:
c_logger.print_eval_step(num_iter, loss_dict,
keep_avg.avg_values)
if args.rank == 0:
# Diagnostic visualizations
idx = np.random.randint(mel_input.shape[0])
const_spec = postnet_output[idx].data.cpu().numpy()
gt_spec = linear_input[idx].data.cpu().numpy() if c.model in [
"Tacotron", "TacotronGST"
] else mel_input[idx].data.cpu().numpy()
align_img = alignments[idx].data.cpu().numpy()
eval_figures = {
"prediction": plot_spectrogram(const_spec,
ap,
output_fig=False),
"ground_truth": plot_spectrogram(gt_spec, ap,
output_fig=False),
"alignment": plot_alignment(align_img, output_fig=False)
}
# Sample audio
if c.model in ["Tacotron", "TacotronGST"]:
eval_audio = ap.inv_spectrogram(const_spec.T)
else:
eval_audio = ap.inv_melspectrogram(const_spec.T)
tb_logger.tb_eval_audios(global_step, {"ValAudio": eval_audio},
c.audio["sample_rate"])
# Plot Validation Stats
if c.bidirectional_decoder or c.double_decoder_consistency:
align_b_img = alignments_backward[idx].data.cpu().numpy()
eval_figures['alignment2'] = plot_alignment(align_b_img,
output_fig=False)
tb_logger.tb_eval_stats(global_step, keep_avg.avg_values)
tb_logger.tb_eval_figures(global_step, eval_figures)
if args.rank == 0 and epoch > c.test_delay_epochs:
if c.test_sentences_file is None:
test_sentences = [
"Unabhängig davon, wer gewinnt, bestehen erhebliche Zweifel, ob die Präsidentschaftswahlen überhaupt verfassungskonform sind.",
]
else:
with open(c.test_sentences_file, "r") as f:
test_sentences = [s.strip() for s in f.readlines()]
# test sentences
test_audios = {}
test_figures = {}
print(" | > Synthesizing test sentences")
speaker_id = 0 if c.use_speaker_embedding else None
style_wav = c.get("gst_style_input")
for idx, test_sentence in enumerate(test_sentences):
try:
wav, alignment, decoder_output, postnet_output, stop_tokens, _ = synthesis(
model,
test_sentence,
c,
use_cuda,
ap,
speaker_id=speaker_id,
style_wav=style_wav,
truncated=False,
enable_eos_bos_chars=c.enable_eos_bos_chars, #pylint: disable=unused-argument
use_griffin_lim=True,
do_trim_silence=False)
file_path = os.path.join(AUDIO_PATH, str(global_step))
os.makedirs(file_path, exist_ok=True)
file_path = os.path.join(file_path,
"TestSentence_{}.wav".format(idx))
ap.save_wav(wav, file_path)
test_audios['{}-audio'.format(idx)] = wav
test_figures['{}-prediction'.format(idx)] = plot_spectrogram(
postnet_output, ap, output_fig=False)
test_figures['{}-alignment'.format(idx)] = plot_alignment(
alignment, output_fig=False)
except: #pylint: disable=bare-except
print(" !! Error creating Test Sentence -", idx)
traceback.print_exc()
tb_logger.tb_test_audios(global_step, test_audios,
c.audio['sample_rate'])
tb_logger.tb_test_figures(global_step, test_figures)
return keep_avg.avg_values
def tts(model,
vocoder_model,
text,
CONFIG,
use_cuda,
ap,
use_gl,
speaker_fileid,
speaker_embedding=None,
gst_style=None):
t_1 = time.time()
waveform, _, _, mel_postnet_spec, _, _ = synthesis(
model,
text,
CONFIG,
use_cuda,
ap,
speaker_fileid,
gst_style,
False,
CONFIG.enable_eos_bos_chars,
use_gl,
speaker_embedding=speaker_embedding)
# grab spectrogram (thx to the nice guys at mozilla discourse for codesnipplet)
if args.save_spectogram:
spec_file_name = args.text.replace(" ", "_")[0:10]
spec_file_name = spec_file_name.translate(
str.maketrans('', '', string.punctuation.replace('_',
''))) + '.npy'
spec_file_name = os.path.join(args.out_path, spec_file_name)
spectrogram = torch.FloatTensor(mel_postnet_spec.T)
spectrogram = spectrogram.unsqueeze(0)
np.save(spec_file_name, spectrogram)
print(" > Saving raw spectogram to " + spec_file_name)
if CONFIG.model == "Tacotron" and not use_gl:
mel_postnet_spec = ap.out_linear_to_mel(mel_postnet_spec.T).T
if not use_gl:
# Use if not computed noise schedule with tune_wavegrad
beta = np.linspace(1e-6, 0.01, 50)
vocoder_model.compute_noise_level(beta)
# Use alternative when using output npy file from tune_wavegrad
# beta = np.load("output-tune-wavegrad.npy", allow_pickle=True).item()
# vocoder_model.compute_noise_level(beta['beta'])
device_type = "cuda" if use_cuda else "cpu"
waveform = vocoder_model.inference(
torch.FloatTensor(mel_postnet_spec.T).to(device_type).unsqueeze(0))
if use_cuda and not use_gl:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
waveform = waveform.squeeze()
rtf = (time.time() - t_1) / (len(waveform) / ap.sample_rate)
tps = (time.time() - t_1) / len(waveform)
print(" > Run-time: {}".format(time.time() - t_1))
print(" > Real-time factor: {}".format(rtf))
print(" > Time per step: {}".format(tps))
return waveform
def tts(
model,
vocoder_model,
text,
CONFIG,
use_cuda,
ap,
use_gl,
speaker_fileid,
speaker_embedding=None,
gst_style=None,
text_is_phonemes=False,
ap_vocoder=None,
scale_factors=None,
):
t_1 = time.time()
waveform, _, _, mel_postnet_spec, _, _ = synthesis(
model=model,
text=text,
CONFIG=CONFIG,
use_cuda=use_cuda,
ap=ap,
speaker_id=speaker_fileid,
style_wav=gst_style,
truncated=False,
enable_eos_bos_chars=CONFIG.enable_eos_bos_chars,
use_griffin_lim=use_gl,
speaker_embedding=speaker_embedding,
backend="torch",
do_trim_silence=False,
text_is_phonemes=text_is_phonemes,
)
if CONFIG.model == "Tacotron" and not use_gl:
mel_postnet_spec = ap.out_linear_to_mel(mel_postnet_spec.T).T
if not use_gl and ap_vocoder:
mel_postnet_spec = ap._denormalize(mel_postnet_spec.T).T
vocoder_input = ap_vocoder._normalize(mel_postnet_spec.T)
if scale_factors:
vocoder_input = torch.nn.functional.interpolate(
torch.tensor(vocoder_input).unsqueeze(0).unsqueeze(0),
scale_factor=scale_factors,
mode="bilinear",
).squeeze(0)
else:
vocoder_input = torch.tensor(vocoder_input).unsqueeze(0)
device_type = "cuda" if use_cuda else "cpu"
waveform = vocoder_model.inference(vocoder_input.to(device_type))
if use_cuda and not use_gl:
waveform = waveform.cpu()
if not use_gl:
waveform = waveform.numpy()
waveform = waveform.squeeze()
rtf = (time.time() - t_1) / (len(waveform) / ap.sample_rate)
tps = (time.time() - t_1) / len(waveform)
print(" > Run-time: {}".format(time.time() - t_1))
print(" > Real-time factor: {}".format(rtf))
print(" > Time per step: {}".format(tps))
return waveform
def evaluate(data_loader, model, criterion, ap, global_step, epoch):
model.eval()
epoch_time = 0
keep_avg = KeepAverage()
c_logger.print_eval_start()
if data_loader is not None:
for num_iter, data in enumerate(data_loader):
start_time = time.time()
# format data
text_input, text_lengths, mel_input, mel_lengths, speaker_c,\
_, _, attn_mask, _ = format_data(data)
# forward pass model
z, logdet, y_mean, y_log_scale, alignments, o_dur_log, o_total_dur = model.forward(
text_input,
text_lengths,
mel_input,
mel_lengths,
attn_mask,
g=speaker_c)
# compute loss
loss_dict = criterion(z, y_mean, y_log_scale, logdet, mel_lengths,
o_dur_log, o_total_dur, text_lengths)
# step time
step_time = time.time() - start_time
epoch_time += step_time
# compute alignment score
align_error = 1 - alignment_diagonal_score(alignments)
loss_dict['align_error'] = align_error
# aggregate losses from processes
if num_gpus > 1:
loss_dict['log_mle'] = reduce_tensor(loss_dict['log_mle'].data,
num_gpus)
loss_dict['loss_dur'] = reduce_tensor(
loss_dict['loss_dur'].data, num_gpus)
loss_dict['loss'] = reduce_tensor(loss_dict['loss'].data,
num_gpus)
# detach loss values
loss_dict_new = dict()
for key, value in loss_dict.items():
if isinstance(value, (int, float)):
loss_dict_new[key] = value
else:
loss_dict_new[key] = value.item()
loss_dict = loss_dict_new
# update avg stats
update_train_values = dict()
for key, value in loss_dict.items():
update_train_values['avg_' + key] = value
keep_avg.update_values(update_train_values)
if c.print_eval:
c_logger.print_eval_step(num_iter, loss_dict,
keep_avg.avg_values)
if args.rank == 0:
# Diagnostic visualizations
# direct pass on model for spec predictions
target_speaker = None if speaker_c is None else speaker_c[:1]
if hasattr(model, 'module'):
spec_pred, *_ = model.module.inference(text_input[:1],
text_lengths[:1],
g=target_speaker)
else:
spec_pred, *_ = model.inference(text_input[:1],
text_lengths[:1],
g=target_speaker)
spec_pred = spec_pred.permute(0, 2, 1)
gt_spec = mel_input.permute(0, 2, 1)
const_spec = spec_pred[0].data.cpu().numpy()
gt_spec = gt_spec[0].data.cpu().numpy()
align_img = alignments[0].data.cpu().numpy()
eval_figures = {
"prediction": plot_spectrogram(const_spec, ap),
"ground_truth": plot_spectrogram(gt_spec, ap),
"alignment": plot_alignment(align_img)
}
# Sample audio
eval_audio = ap.inv_melspectrogram(const_spec.T)
tb_logger.tb_eval_audios(global_step, {"ValAudio": eval_audio},
c.audio["sample_rate"])
# Plot Validation Stats
tb_logger.tb_eval_stats(global_step, keep_avg.avg_values)
tb_logger.tb_eval_figures(global_step, eval_figures)
if args.rank == 0 and epoch >= c.test_delay_epochs:
if c.test_sentences_file is None:
test_sentences = [
"It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
"Be a voice, not an echo.",
"I'm sorry Dave. I'm afraid I can't do that.",
"This cake is great. It's so delicious and moist.",
"Prior to November 22, 1963."
]
else:
with open(c.test_sentences_file, "r") as f:
test_sentences = [s.strip() for s in f.readlines()]
# test sentences
test_audios = {}
test_figures = {}
print(" | > Synthesizing test sentences")
if c.use_speaker_embedding:
if c.use_external_speaker_embedding_file:
speaker_embedding = speaker_mapping[list(
speaker_mapping.keys())[randrange(
len(speaker_mapping) - 1)]]['embedding']
speaker_id = None
else:
speaker_id = 0
speaker_embedding = None
else:
speaker_id = None
speaker_embedding = None
style_wav = c.get("style_wav_for_test")
for idx, test_sentence in enumerate(test_sentences):
try:
wav, alignment, _, postnet_output, _, _ = synthesis(
model,
test_sentence,
c,
use_cuda,
ap,
speaker_id=speaker_id,
speaker_embedding=speaker_embedding,
style_wav=style_wav,
truncated=False,
enable_eos_bos_chars=c.enable_eos_bos_chars, #pylint: disable=unused-argument
use_griffin_lim=True,
do_trim_silence=False)
file_path = os.path.join(AUDIO_PATH, str(global_step))
os.makedirs(file_path, exist_ok=True)
file_path = os.path.join(file_path,
"TestSentence_{}.wav".format(idx))
ap.save_wav(wav, file_path)
test_audios['{}-audio'.format(idx)] = wav
test_figures['{}-prediction'.format(idx)] = plot_spectrogram(
postnet_output, ap)
test_figures['{}-alignment'.format(idx)] = plot_alignment(
alignment)
except: #pylint: disable=bare-except
print(" !! Error creating Test Sentence -", idx)
traceback.print_exc()
tb_logger.tb_test_audios(global_step, test_audios,
c.audio['sample_rate'])
tb_logger.tb_test_figures(global_step, test_figures)
return keep_avg.avg_values
def synthesize(self, response_text):
# pre cleaning
text = self.pre_clean(response_text)
if self.model_name == 'nvidia':
# TODO choose language?
cleaner = ['catalan_cleaners']
# Prepare text input
sequence = np.array(text_to_sequence(text, cleaner))[None, :]
sequence = torch.from_numpy(sequence).to(device='cpu', dtype=torch.int64)
# TODO run within the queue
# decode text input
mel_outputs, mel_outputs_postnet, _, alignments = self.model.inference(sequence)
# TODO run within the queue
# Synthesize using neurips Melgan
with torch.no_grad():
audio = self.vocoder.inverse(mel_outputs_postnet.float())
audio_numpy = audio[0].data.cpu().numpy()
# normalize and convert from float32 to int16 pcm
audio_numpy /= np.max(np.abs(audio_numpy))
audio_numpy *= 32768*0.99
waveform = audio_numpy.astype(np.int16)
elif self.model_name == 'coqui':
wavs = []
sens = self.split_into_sentences(text)
for sen in sens:
outputs = synthesis(model=self.tts_model,
text=text,
CONFIG=self.tts_config,
use_cuda=False,
ap=self.ap,
speaker_id=None,
style_wav=None,
enable_eos_bos_chars=self.tts_config.enable_eos_bos_chars,
use_griffin_lim=False,
d_vector=None)
# extract and normalize the spectogram
mel_postnet_spec = outputs["outputs"]["model_outputs"][0].detach().cpu().numpy()
mel_postnet_spec = self.ap.denormalize(mel_postnet_spec.T).T
vocoder_input = self.vocoder_ap.normalize(mel_postnet_spec.T)
vocoder_input = torch.tensor(vocoder_input).unsqueeze(0)
# Synthesize using neurips Melgan
audio = self.vocoder.inverse(vocoder_input.to('cpu'))
audio_numpy = audio[0].data.cpu().numpy()
audio_numpy /= np.max(np.abs(audio_numpy))
audio_numpy *= 32768*0.99
waveform = list(audio_numpy.astype(np.int16).squeeze())
wavs += waveform
waveform = audio_numpy.astype(np.int16).squeeze()
else:
raise ValueError('% unknown model name for synthesis'%self.model_name)
# out
out = io.BytesIO()
# save
scipy.io.wavfile.write(out, 22050, waveform)
return out.getvalue()
