def tts(model, text):
"""Convert text to speech waveform given a Tacotron model.
"""
if use_cuda:
model = model.cuda()
# TODO: Turning off dropout of decoder's prenet causes serious performance
# regression, not sure why.
# model.decoder.eval()
model.encoder.eval()
model.postnet.eval()
sequence = np.array(text_to_sequence(text, [hparams.cleaners]))
sequence = Variable(torch.from_numpy(sequence)).unsqueeze(0)
if use_cuda:
sequence = sequence.cuda()
# Greedy decoding
mel_outputs, linear_outputs, alignments = model(sequence)
linear_output = linear_outputs[0].cpu().data.numpy()
spectrogram = audio._denormalize(linear_output)
alignment = alignments[0].cpu().data.numpy()
# Predicted audio signal
waveform = audio.inv_spectrogram(linear_output.T)
return waveform, alignment, spectrogram
def save_spectrogram(path, linear_output):
spectrogram = audio._denormalize(linear_output)
plt.figure(figsize=(16, 10))
plt.imshow(spectrogram.T, aspect="auto", origin="lower")
plt.colorbar()
plt.tight_layout()
plt.savefig(path, format="png")
plt.close()