def synthesis_griffin_lim(text_seq,
model,
alpha=1.0,
mode="",
num=100,
check=True):
text_seq = text_seq[:-1]
text = text_to_sequence(text_seq, hp.hparams.text_cleaners)
text = text + [0]
text = np.stack([np.array(text)])
text = torch.from_numpy(text).long().to(device)
sequence = np.array(text_to_sequence(text_seq,
hp.hparams.text_cleaners))[None, 1]
pos = torch.stack([torch.Tensor([i + 1 for i in range(text.size(1))])])
pos = pos.long().to(device)
model.eval()
with torch.no_grad():
mel, mel_postnet = model(text, pos, alpha=alpha)
if not os.path.exists("results_kor_0730_nam_95000"):
os.mkdir("results_kor_0730_nam_95000")
new_name = text_seq.replace(" ", "_")
new_name = new_name.replace("?", "_")
new_name = new_name[:-1]
new_name2 = new_name + str(num) + mode + ".wav"
new_name3 = "results_kor_0730_nam_95000/" + new_name2
mel = mel[0].cpu().numpy().T
mel_postnet = mel_postnet[0].cpu().numpy().T
plot_data([mel, mel_postnet], file_name=new_name)
start = int(round(time.time() * 1000))
wav = audio.inv_mel_spectrogram(mel_postnet)
end = int(round(time.time() * 1000))
audio.save_wav(wav, os.path.join("results_kor_0730_nam_95000", new_name2))
clean_text = new_name.replace("_", " ")
if check:
x, _, _, y, _, _ = WERCER([new_name3], [str(clean_text)])
else:
x = 0
y = 0
print("Total time : ", end - start)
print()
return new_name, x, y
def synthesis_griffin_lim(text_seq, model, alpha=1.0, mode="", num=100):
text = text_to_sequence(text_seq, hp.hparams.text_cleaners)
text = text + [0]
text = np.stack([np.array(text)])
text = torch.from_numpy(text).long().to(device)
pos = torch.stack([torch.Tensor([i + 1 for i in range(text.size(1))])])
pos = pos.long().to(device)
start = int(round(time.time() * 1000))
model.eval()
with torch.no_grad():
mel, mel_postnet = model(text, pos, alpha=alpha)
end = int(round(time.time() * 1000))
tt = end - start
print("Total - making mel : %d ms\n" % tt)
mel = mel[0].cpu().numpy().T
mel_postnet = mel_postnet[0].cpu().numpy().T
#plot_data([mel, mel_postnet])
wav = audio.inv_mel_spectrogram(mel_postnet)
print("Wav Have Been Synthesized.\n")
if not os.path.exists("results"):
os.mkdir("results")
new_name = text_seq.replace(" ", "_")
audio.save_wav(
wav, os.path.join("results", new_name + str(num) + mode + ".wav"))
return new_name
def get_tacotron2_alignment_test(text_seq):
hparams = hp_tacotron2.create_hparams()
hparams.sampling_rate = hp.sample_rate
checkpoint_path = os.path.join("Tacotron2",
os.path.join("outdir", "checkpoint_51000"))
tacotron2 = train_tacotron2.load_model(hparams)
tacotron2.load_state_dict(torch.load(checkpoint_path)["state_dict"])
_ = tacotron2.cuda().eval().half()
sequence = np.array(text_to_sequence(text_seq,
hp.hparams.text_cleaners))[None, :]
print("sequence size", np.shape(sequence))
sequence = torch.autograd.Variable(
torch.from_numpy(sequence)).cuda().long()
mel, mel_postnet, _, alignment = tacotron2.inference(sequence)
wav = audio.inv_mel_spectrogram(mel_postnet.float().data.cpu().numpy()[0])
file_name = text_seq.replace(" ", "_")
audio.save_wav(wav, "%s.wav" % file_name)
alignment = alignment.float().data.cpu().numpy()[0]
print("alignment size", np.shape(alignment))
get_D(alignment)
return alignment
def __getitem__(self, index):
id = self.meta_data.iloc[index]['id']
text = self.meta_data.iloc[index]['text']
input_seq = text_to_sequence(text)
mels = np.load(f'{self.path}/mels/{id}')
linears = np.load(f'{self.path}/linears/{id}')
return input_seq, (mels, linears)
def synthesis_waveglow(text_seq, model, waveglow, alpha=1.0, mode=""):
text = text_to_sequence(text_seq, hp.hparams.text_cleaners)
text = text + [0]
text = np.stack([np.array(text)])
text = torch.from_numpy(text).long().to(device)
pos = torch.stack([torch.Tensor([i + 1 for i in range(text.size(1))])])
pos = pos.long().to(device)
model.eval()
with torch.no_grad():
_, mel_postnet = model(text, pos, alpha=alpha)
with torch.no_grad():
wav = waveglow.infer(mel_postnet, sigma=0.666)
print("Wav Have Been Synthesized.")
if not os.path.exists("results"):
os.mkdir("results")
audio.save_wav(wav[0].data.cpu().numpy(),
os.path.join("results", text_seq + mode + ".wav"))
def synthesis_griffin_lim(text_seq, model, alpha=1.0, mode=""):
text = text_to_sequence(text_seq, hp.hparams.text_cleaners)
text = text + [0]
text = np.stack([np.array(text)])
text = torch.from_numpy(text).long().to(device)
pos = torch.stack([torch.Tensor([i + 1 for i in range(text.size(1))])])
pos = pos.long().to(device)
model.eval()
with torch.no_grad():
mel, mel_postnet = model(text, pos, alpha=alpha)
mel = mel[0].cpu().numpy().T
mel_postnet = mel_postnet[0].cpu().numpy().T
plot_data([mel, mel_postnet])
wav = audio.inv_mel_spectrogram(mel_postnet)
print("Wav Have Been Synthesized.")
if not os.path.exists("results"):
os.mkdir("results")
audio.save_wav(wav, os.path.join("results", text_seq + mode + ".wav"))
def __getitem__(self, idx):
index = idx + 1
mel_name = os.path.join(self.dataset_path, self.paths[idx])
#mel_name = os.path.join(
#t self.dataset_path, "nam_speech-mel-%05d.npy" % index)
# mel_name = os.path.join(
#self.dataset_path, "nam_speech-mel-%05d.npy" % index)
# self.dataset_path, "ljspeech-mel-%05d.npy" %index)
# print(mel_name)
mel_np = np.load(mel_name)
# print(self.text[idx])
# print(mel_name)
character = self.text[idx]
character = text_to_sequence(character, hp.hparams.text_cleaners)
character = np.array(character)
if not hp.pre_target:
return {"text": character, "mel": mel_np}
else:
# filename = "0"*(7-len(str(idx)))+str(idx)
#i=0
# while(i<13):
align_path = os.path.join(hp.alignment_target_path,
self.path_as[idx])
# if os.path.exists(path):
# print(align_path)
alignment = np.load(align_path)
#print("text %s" %character)
#print("alignment %s" %alignment)
#print("mel %s" %mel_np)
#print(mel_name)
#print(align_path)
return {"text": character, "mel": mel_np, "alignment": alignment}
def synthesis_griffin_lim(text_seq,
model,
alpha=1.0,
mode="",
num=100,
check=True,
cute=False):
text = text_to_sequence(text_seq, hp.hparams.text_cleaners)
text = text + [0]
text = np.stack([np.array(text)])
text = torch.from_numpy(text).cuda().to(device)
pos = torch.stack([torch.Tensor([i + 1 for i in range(text.size(1))])])
pos = pos.long().to(device)
model.eval()
#mel = generate_mels(model, text, pos, 1, 0)
with torch.no_grad():
mel, mel_postnet = model(text, pos, alpha=alpha)
if not os.path.exists("results_kor_0730_indiv"):
os.mkdir("results_kor_0730_indiv")
new_name = text_seq.replace(" ", "_")
new_name = new_name.replace("?", "_")
if (cute):
new_name2 = new_name + "_cute"
new_name2 = new_name + str(num) + ".wav"
new_name3 = "results_kor_0730_indiv/" + new_name2
if (cute):
#high-pitched sound
mel_postnet = mel_postnet[0].cpu().numpy().T
else:
#print('mel', mel.max(), mel.mean(), mel.min())
#print('mel.shape' , mel_postnet.shape)
mel_postnet = mel_postnet.data.cpu().numpy()[0].T
mel_postnet = mel_postnet[:, :-1]
mel_postnet = np.append(mel_postnet,
np.ones((80, 0), dtype=np.float32) * -4.0,
axis=1)
#print(mel.shape)
mel = mel[0].cpu().numpy().T
#print('mel_postnet', mel_postnet.max(), mel_postnet.mean(), mel_postnet.min())
plot_data([mel, mel_postnet], file_name=new_name)
mels = []
mels.append(mel_postnet)
if (cute):
wav = audio.inv_mel_spectrogram(mel_postnet)
else:
stft = audio.taco_stft()
wav = mels_to_wavs_GL(mels, stft)
audio.save_wav(
wav,
os.path.join("results_kor_0730_indiv", new_name + str(num) + ".wav"))
clean_text = new_name.replace("_", " ")
if check:
x, _, _, y, _, _ = WERCER([new_name3], [str(clean_text)])
else:
x = 0
y = 0
return new_name, x, y
def get_text(self, text):
text_norm = torch.IntTensor(text_to_sequence(text, self.text_cleaners))
return text_norm
break
return txt
if __name__ == "__main__":
# Test
tacotron2 = get_tacotron2()
text_path = os.path.join('dataset/nam', "train.txt")
text = process_text(text_path)
i = 0
for i in range(len(text)):
text_seq = np.array(text_to_sequence(text[i],
['korean_cleaners']))[None, :]
text_seq = torch.from_numpy(text_seq)
alignment = get_one_alignment(text_seq, tacotron2)
file_name = "%d.npy" % i
dir_path = "./alignment_targets_nam"
file_path = os.path.join(dir_path, file_name)
if not os.path.exists(dir_path):
os.mkdir(dir_path)
if not os.path.exists(file_path):
f = open(file_path, 'a+')
f.close()
np.save(file_path, alignment)
if i % 100 == 0:
print("current step : %d\n" % i)
