def test_compute_monophone_ppg(self):
deps = ppg.DependenciesPPG()
wave_data = feat.read_wav_kaldi(self.wav_path)
ppgs = ppg.compute_monophone_ppg(wave_data, deps.nnet, deps.lda,
deps.monophone_trans)
reduce_ppg_dim = deps.monophone_trans.num_rows
self.assertEqual(ppgs.shape[1], reduce_ppg_dim)
self.assertAlmostEqual(ppgs.sum(), ppgs.shape[0], 1)
def image_ppg(ppg_np):
"""
Input:
ppg: numpy array
Return:
ax: 画布信息
im：图像信息
"""
ppg_deps = ppg.DependenciesPPG()
ppg_M = Matrix(ppg_np)
monophone_ppgs = ppg.reduce_ppg_dim(ppg_M, ppg_deps.monophone_trans)
monophone_ppgs = monophone_ppgs.numpy().T
fig, ax = plt.subplots(figsize=(10, 6))
im = ax.imshow(monophone_ppgs,
aspect="auto",
origin="lower",
interpolation='none')
return ax, im
def get_monophone_ppg(self) -> ndarray:
"""A wrapper function to initialize the monophone ppg of this utterance.
Requires non-empty waveform, fs, and kaldi_shift.
Returns:
The monophone ppgs in numpy ndarray format.
"""
if self.kaldi_shift < 1: # ms
raise ValueError('Invalid frame kaldi frame shift parameter %d.',
self.kaldi_shift)
if self.wav.size == 0 or self.fs < 0:
raise ValueError('To perform alignment, the object must contain '
'valid speech data and sampling frequency.')
wav_kaldi = read_wav_kaldi_internal(self.wav, self.fs)
ppg_deps = ppg.DependenciesPPG()
self.monophone_ppg = ppg.compute_monophone_ppg(wav_kaldi, ppg_deps.nnet,
ppg_deps.lda,
ppg_deps.monophone_trans,
self.kaldi_shift)
return self.monophone_ppg
logging.debug('Denoiser strength: %f', denoiser_strength)
logging.debug('Denoiser mode: %s', denoiser_mode)
hparams = create_hparams_stage()
taco_stft = TacotronSTFT(hparams.filter_length, hparams.hop_length,
hparams.win_length, hparams.n_acoustic_feat_dims,
hparams.sampling_rate, hparams.mel_fmin,
hparams.mel_fmax)
# Load models.
tacotron_model = load_model(hparams)
tacotron_model.load_state_dict(torch.load(checkpoint_path)['state_dict'])
_ = tacotron_model.eval()
waveglow_model = load_waveglow_model(waveglow_path)
deps = ppg.DependenciesPPG()
if os.path.isfile(teacher_utt_path):
logging.info('Perform AC on %s', teacher_utt_path)
teacher_ppg = get_ppg(teacher_utt_path, deps)
ac_mel = get_inference(teacher_ppg, tacotron_model, is_clip)
ac_wav = waveglow_audio(ac_mel, waveglow_model, waveglow_sigma, True)
ac_wav = denoiser(ac_wav,
strength=denoiser_strength)[:, 0].cpu().numpy().T
output_file = os.path.join(output_dir, 'ac.wav')
wavfile.write(output_file, fs, ac_wav)
else:
logging.warning('Missing %s', teacher_utt_path)
logging.info('Done!')
def test_ppg_dependencies(self):
deps = ppg.DependenciesPPG()
self.assertIsNotNone(deps)