def generate(model, text):
# Text to index sequence
cleaner_names = [x.strip() for x in hp.cleaners.split(',')]
seq = np.expand_dims(np.asarray(text_to_sequence(text, cleaner_names),
dtype=np.int32),
axis=0)
# Provide [GO] Frame
mel_input = np.zeros([seq.shape[0], hp.num_mels, 1], dtype=np.float32)
# Variables
characters = Variable(torch.from_numpy(seq).type(torch.cuda.LongTensor),
volatile=True).cuda()
mel_input = Variable(torch.from_numpy(mel_input).type(
torch.cuda.FloatTensor),
volatile=True).cuda()
# Spectrogram to wav
_, linear_output = model.forward(characters, mel_input)
wav = inv_spectrogram(linear_output[0].data.cpu().numpy())
wav = wav[:find_endpoint(wav)]
out = io.BytesIO()
save_wav(wav, out)
return out.getvalue()
def generate(model, text, device, writer, curr, _tt):
# Text to index sequence
cleaner_names = [x.strip() for x in hp.cleaners.split(',')]
seq = np.expand_dims(np.asarray(text_to_sequence(text), dtype=np.int32),
axis=0)
# Provide [GO] Frame
mel_input = np.zeros([seq.shape[0], hp.num_mels, 1], dtype=np.float32)
# Variables
characters = torch.from_numpy(seq).type(torch.cuda.LongTensor).to(device)
mel_input = torch.from_numpy(mel_input).type(
torch.cuda.FloatTensor).to(device)
mel_input = torch.transpose(mel_input, 1, 2)
# Spectrogram to wav
_, linear_output = model(characters, mel_input, False)
linear_output = torch.transpose(linear_output, 1, 2)
wav = inv_spectrogram(linear_output[0].data.cpu().numpy())
wav = wav[:find_endpoint(wav)]
wav_tensor = wav * 1.0 / max(0.01, np.max(np.abs(wav)))
wav_tensor = torch.Tensor(wav).to(device).view(1, -1)
writer.add_audio('audio_result_%02d' % (_tt), wav_tensor, curr,
hp.sample_rate)
out = io.BytesIO()
save_wav(wav, out)
return out.getvalue()
def generate_audio(self):
print('Generating Audio Samples')
# Text to index sequence
characters = []
for text in self.sentences:
text = np.asarray(text_to_sequence(text, [hp.cleaners]),
dtype=np.int32)
characters.append(text)
characters = _prepare_data(characters).astype(np.int32)
characters = torch.from_numpy(characters).long().to(self.args.device)
# Provide [GO] Frame
mel_input = torch.zeros([characters.shape[0], hp.num_mels, 1],
dtype=torch.float).to(self.args.device)
print('char: ', characters.shape)
print('mel input: ', mel_input.shape)
self.model.eval()
# Spectrogram to wav
_, linear_output = self.model(characters, mel_input)
for i in range(linear_output.shape[0]):
wav = inv_spectrogram(linear_output[i].data.cpu().numpy())
wav = wav[:find_endpoint(wav)].astype(np.float32)
print('wav: ', wav.shape)
print('wav max: ', wav.max())
print('wav min: ', wav.min())
print('wav: ', wav.dtype)
self.writer.add_audio('audio', wav, self.epoch, sample_rate=16000)
def main(args):
device = torch.device('cuda:0')
if 'english' in hp.cleaners:
_symbols = en_symbols
elif 'korean' in hp.cleaners:
_symbols = symbols
model = Tacotron(len(_symbols)).to(device)
checkpoint = torch.load(args.checkpoint_path)
model.load_state_dict(checkpoint['model'])
model = model.eval()
sentences = [
'Scientists at the CERN laboratory say they have discovered a new particle.',
'President Trump met with other leaders at the Group of 20 conference.',
'Generative adversarial network or variational auto-encoder.',
'Does the quick brown fox jump over the lazy dog?'
]
# Text to index sequence
for i, ele in enumerate(sentences):
cleaner_names = [x.strip() for x in hp.cleaners.split(',')]
seq = np.expand_dims(np.asarray(text_to_sequence(ele), dtype=np.int32),
axis=0)
# Provide [GO] Frame
mel_input = np.zeros([seq.shape[0], hp.num_mels, 1], dtype=np.float32)
# Variables
characters = torch.from_numpy(seq).type(
torch.cuda.LongTensor).to(device)
mel_input = torch.from_numpy(mel_input).type(
torch.cuda.FloatTensor).to(device)
mel_input = torch.transpose(mel_input, 1, 2)
# Spectrogram to wav
mel_output, linear_output = model(characters, mel_input, False)
linear_output = torch.transpose(linear_output, 1, 2)
wav = inv_spectrogram(linear_output[0].data.cpu().numpy())
_wav = wav[:find_endpoint(wav)]
out = io.BytesIO()
save_wav(_wav, out)
tt = out.getvalue()
f = open('./result_%02d.wav' % i, 'wb')
f.write(tt)
f.close()