def preprocess(self):
wav = self.ui.selected_utterance.wav
out = aukit.remove_noise(wav, sr=Synthesizer.sample_rate)
hp = aukit.Dict2Obj({})
hp["vad_window_length"] = 10 # milliseconds
hp["vad_moving_average_width"] = 2
hp["vad_max_silence_length"] = 2
hp["audio_norm_target_dBFS"] = -32
hp["sample_rate"] = 16000
hp["int16_max"] = (2**15) - 1
out = trim_long_silences(out, hparams=hp)
spec = Synthesizer.make_spectrogram(out)
self.ui.draw_align(spec[::-1], "current")
def synthesize(self):
self.ui.log("Generating the mel spectrogram...")
self.ui.set_loading(1)
# Synthesize the spectrogram
if self.synthesizer is None:
model_dir = Path(self.ui.current_synthesizer_model_dir)
checkpoints_dir = model_dir.joinpath("checkpoints")
hp_path = model_dir.joinpath(
"metas", "hparams.json") # load from trained models
if hp_path.exists():
hparams = aukit.Dict2Obj(
json.load(open(hp_path, encoding="utf8")))
else:
hparams = None
self.synthesizer = Synthesizer(checkpoints_dir,
low_mem=self.low_mem,
hparams=hparams)
if not self.synthesizer.is_loaded():
self.ui.log("Loading the synthesizer %s" %
self.synthesizer.checkpoint_fpath)
ptext = self.ui.text_prompt.toPlainText()
texts = ptext.split("\n")
embed = self.ui.selected_utterance.embed
embeds = np.stack([embed] * len(texts))
specs, aligns = self.synthesizer.synthesize_spectrograms(
texts, embeds, return_alignments=True)
breaks = [spec.shape[1] for spec in specs]
spec = np.concatenate(specs, axis=1)
align = np.concatenate(aligns, axis=1)
fref = self.ui.selected_utterance.name
ftext = '。'.join(texts)
ftime = '{}'.format(time_formatter())
fname = filename_formatter('{}_{}_{}zi_{}.npy'.format(
fref, ftime, len(ftext), ftext))
np.save(self._out_mel_dir.joinpath(fname), spec,
allow_pickle=False) # save
self.ui.draw_spec(spec, "generated")
self.ui.draw_align(align, "generated")
self.current_generated = (self.ui.selected_utterance.speaker_name,
spec, breaks, None)
self.ui.set_loading(0)
print(
"Found %d GPUs available. Using GPU %d (%s) of compute capability %d.%d with "
"%.1fGb total memory.\n" %
(torch.cuda.device_count(), device_id, gpu_properties.name,
gpu_properties.major, gpu_properties.minor,
gpu_properties.total_memory / 1e9))
## Load the models one by one.
print("Preparing the encoder, the synthesizer and the vocoder...")
encoder.load_model(args.enc_model_fpath, device='cpu')
# 从模型目录导入hparams
hp_path = args.syn_model_dir.parent.joinpath(
"metas", "hparams.json") # load from trained models
if hp_path.exists():
hparams = aukit.Dict2Obj(json.load(open(hp_path, encoding="utf8")))
print('hparams:')
print(
json.dumps({k: v
for k, v in hparams.items()},
ensure_ascii=False,
indent=4))
else:
hparams = None
print('hparams:', hparams)
synthesizer = Synthesizer(args.syn_model_dir,
low_mem=args.low_mem,
hparams=hparams)
# vocoder.load_model(args.voc_model_fpath)