def say(sentence):
new_sentence=" " .join([num2words(w) if w.isdigit() else w for w in sentence.split()])
normalized_sentence = "".join([c if c.lower() in vocab else '' for c in new_sentence])
print(normalized_sentence)
sentences = [normalized_sentence]
max_N = len(normalized_sentence)
L = torch.from_numpy(get_test_data(sentences, max_N))
zeros = torch.from_numpy(np.zeros((1, hp.n_mels, 1), np.float32))
Y = zeros
A = None
for t in range(hp.max_T):
_, Y_t, A = text2mel(L, Y, monotonic_attention=True)
Y = torch.cat((zeros, Y_t), -1)
_, attention = torch.max(A[0, :, -1], 0)
attention = attention.item()
if L[0, attention] == vocab.index('E'): # EOS
break
_, Z = ssrn(Y)
i=int(0)
Z = Z.cpu().detach().numpy()
save_to_wav(Z[0, :, :].T, '%d.wav' % (i + 1))
playsound('1.wav')
sys.exit(1)
# synthetize by one by one because there is a batch processing bug!
for i in range(len(SENTENCES)):
sentences = [SENTENCES[i]]
max_N = len(SENTENCES[i])
L = torch.from_numpy(get_test_data(sentences, max_N))
zeros = torch.from_numpy(np.zeros((1, hp.n_mels, 1), np.float32))
Y = zeros
A = None
for t in tqdm(range(hp.max_T)):
_, Y_t, A = text2mel(L, Y, monotonic_attention=True)
Y = torch.cat((zeros, Y_t), -1)
_, attention = torch.max(A[0, :, -1], 0)
attention = attention.item()
if L[0, attention] == vocab.index('E'): # EOS
break
_, Z = ssrn(Y)
Y = Y.cpu().detach().numpy()
A = A.cpu().detach().numpy()
Z = Z.cpu().detach().numpy()
save_to_png('samples/%d-att.png' % (i + 1), A[0, :, :])
save_to_png('samples/%d-mel.png' % (i + 1), Y[0, :, :])
save_to_png('samples/%d-mag.png' % (i + 1), Z[0, :, :])
save_to_wav(Z[0, :, :].T, 'samples/%d-wav.wav' % (i + 1))
# text2mel = text2mel.eval()
for sentence in SENTENCES:
with torch.no_grad():
L = torch.from_numpy(get_test_data(sentence)).to(device)
zeros = torch.from_numpy(np.zeros((1, hp.n_mels, 1), np.float32)).to(device)
Y = zeros
# A = None
while True:
_, Y_t, A = text2mel(L, Y, monotonic_attention=True)
Y = torch.cat((zeros, Y_t), -1)
_, attention = torch.max(A[0, :, -1], 0)
attention = attention.item()
if L[0, attention] == vocab.index('E'): # EOS
print(f'{sentence} ok!')
break
_, Z = ssrn(Y)
# Y = Y.cpu().detach().numpy()
# A = A.cpu().detach().numpy()
Z = Z.cpu().detach().numpy()
if not os.path.isdir(f'samples/{folder}'):
os.mkdir(f'samples/{folder}')
if not os.path.isdir(f'samples/{folder}/{filename}'):
os.mkdir(f'samples/{folder}/{filename}')
# save_to_png('samples/%d-att.png' % (i + 1), A[0, :, :])
# save_to_png('samples/%d-mel.png' % (i + 1), Y[0, :, :])
# save_to_png('samples/%d-mag.png' % (i + 1), Z[0, :, :])
save_to_wav(Z[0, :, :].T, f'samples/{folder}/{filename}/{sentence}.wav')
for t in tqdm(range(hp.max_T)):
_, Y_t, A = text2mel(L, Y, speakers, monotonic_attention=True)
Y = torch.cat((zeros, Y_t), -1)
_, attention = torch.max(A[0, :, -1], 0)
attention = attention.item()
if L[0, attention] == vocab.index('E'): # EOS
break
_, Z = ssrn(Y)
Y = Y.cpu().detach().numpy()
A = A.cpu().detach().numpy()
Z = Z.cpu().detach().numpy()
save_to_png('samples/%d-att.png' % (i + 1), A[0, :, :])
save_to_png('samples/%d-mel.png' % (i + 1), Y[0, :, :])
save_to_png('samples/%d-mag.png' % (i + 1), Z[0, :, :])
# import matplotlib.pyplot as plt
# a = self.embeddings(torch.tensor([x for x in range(10)]).cuda())
# fig, ax = plt.subplots()
# plt.imshow(A[0,:,:23])
# labels = [item.get_text() for item in ax.get_xticklabels()]
# labels=[x for x in sentence]
#
# ax.set_yticklabels(labels[::-1])
# plt.show()
# save_to_wav(Z[0, :, :].T, 'samples/%d-wav.wav' % (i + 1))
print('saving for speaker: ', speaker)
save_to_wav(Z[0, :, :].T, 'samples/%d-%s-wav.wav' % ((i + 1), speaker))
sentences = [SENTENCES[i]]
max_N = len(SENTENCES[i])
L = torch.from_numpy(get_test_data(sentences, max_N))
zeros = torch.from_numpy(np.zeros((1, hp.n_mels, 1), np.float32))
Y = zeros
A = None
for t in tqdm(range(hp.max_T)):
_, Y_t, A = text2mel(L, Y, monotonic_attention=True)
Y = torch.cat((zeros, Y_t), -1)
_, attention = torch.max(A[0, :, -1], 0)
attention = attention.item()
if L[0, attention] == vocab.index('E'): # EOS
break
_, Z = ssrn(Y)
Y = Y.cpu().detach().numpy()
A = A.cpu().detach().numpy()
Z = Z.cpu().detach().numpy()
save_to_png('samples/samples_bea_sleepiness_3k/%d-att.png' % (i + 1),
A[0, :, :])
save_to_png('samples/samples_bea_sleepiness_3k/%d-mel.png' % (i + 1),
Y[0, :, :])
save_to_png('samples/samples_bea_sleepiness_3k/%d-mag.png' % (i + 1),
Z[0, :, :])
save_to_wav(Z[0, :, :].T,
'samples/samples_bea_sleepiness_3k/%d-wav.wav' % (i + 1))
max_N = len(SENTENCES[i])
L = torch.from_numpy(get_test_data(sentences, max_N))
zeros = torch.from_numpy(np.zeros((1, hp.n_mels, 1), np.float32))
Y = zeros
A = None
for t in tqdm(range(hp.max_T)):
_, Y_t, A = text2mel(L, Y, monotonic_attention=True)
Y = torch.cat((zeros, Y_t), -1)
_, attention = torch.max(A[0, :, -1], 0)
attention = attention.item()
if L[0, attention] == vocab.index('E'): # EOS
break
_, Z = ssrn(Y)
Y = Y.cpu().detach().numpy()
A = A.cpu().detach().numpy()
Z = Z.cpu().detach().numpy()
#print("Z", Z[0, :, :])
save_dir = hp.synthesize_samples_dir
file_name = save_dir + '/' + t2mel_step_str + "_" + ssrn_step_str
os.makedirs(file_name, exist_ok=True)
save_to_png(file_name + '/%d-att.png' % (i + 1), A[0, :, :])
save_to_png(file_name + '/%d-mel.png' % (i + 1), Y[0, :, :])
save_to_png(file_name + '/%d-mag.png' % (i + 1), Z[0, :, :])
save_to_wav(Z[0, :, :].T, file_name + '/%d-wav.wav' % (i + 1))