def tts_predict(model, ppg, id_speaker):
# 准备输入的数据并转换到GPU
ppg = Variable(torch.from_numpy(ppg)).unsqueeze(0).float()
id_speaker = torch.LongTensor([id_speaker])
print('orig:', id_speaker)
print(id_speaker.shape)
# id_speaker = id_speaker.unsqueeze(0)
print(ppg.size())
print(ppg.shape)
print(ppg.type())
print('---------- id_speaker')
print(id_speaker.size())
print(id_speaker.shape)
print(id_speaker.type())
print(id_speaker)
if use_cuda:
ppg = ppg.cuda()
id_speaker = id_speaker.cuda()
# 进行预测并数据转换到CPU
mel_pred, spec_pred = model(ppg, id_speaker)
mel_pred = mel_pred[0].cpu().data.numpy()
spec_pred = spec_pred[0].cpu().data.numpy()
# vocoder合成音频波形文件
mel_pred_audio = normalized_db_mel2wav(mel_pred)
spec_pred_audio = normalized_db_spec2wav(spec_pred)
return mel_pred, spec_pred, mel_pred_audio, spec_pred_audio
def main():
#这一部分用于处理LJSpeech格式的数据集
a = open(meta_path, 'r').readlines()
b = []
i = 0
while i < len(a):
t = a[i][0:6]
b.append(t)
i += 2
print(b[:2])
a = b
# a = [i.strip().split('|')[0] for i in a]
cnt = 0
cnt_list = []
bad_cnt = 0
bad_list = []
for fname in tqdm(a):
try:
# 提取声学参数
wav_f = os.path.join(wav_dir, fname + '.wav')
wav_arr = load_wav(wav_f)
mfcc_feats = wav2unnormalized_mfcc(wav_arr)
mel_feats = wav2normalized_db_mel(wav_arr)
spec_feats = wav2normalized_db_spec(wav_arr)
# 验证声学参数提取的对
save_name = fname + '.npy'
save_mel_rec_name = fname + '_mel_rec.wav'
save_spec_rec_name = fname + '_spec_rec.wav'
# 这句话有可能错，不知道为什么，可能是服务器临时变动有关
ppg_already_feats = np.load(os.path.join(ppg_dir, save_name))
assert ppg_already_feats.shape[0] == mfcc_feats.shape[0]
assert mfcc_feats.shape[0] == mel_feats.shape[0] and mel_feats.shape[0] == spec_feats.shape[0]
write_wav(os.path.join(rec_wav_dir, save_mel_rec_name), normalized_db_mel2wav(mel_feats))
write_wav(os.path.join(rec_wav_dir, save_spec_rec_name), normalized_db_spec2wav(spec_feats))
# 存储声学参数
mfcc_save_name = os.path.join(mfcc_dir, save_name)
mel_save_name = os.path.join(mel_dir, save_name)
spec_save_name = os.path.join(spec_dir, save_name)
np.save(mfcc_save_name, mfcc_feats)
np.save(mel_save_name, mel_feats)
np.save(spec_save_name, spec_feats)
f_good_meta.write(fname + '\n')
cnt_list.append(fname)
cnt += 1
except:
bad_list.append(fname)
bad_cnt += 1
# print(cnt)
# break
print(cnt)
print('bad:', bad_cnt)
print(bad_list)
return
def tts_predict(model, ppg):
# 准备输入的数据并转换到GPU
ppg = Variable(torch.from_numpy(ppg)).unsqueeze(0).float()
if use_cuda:
ppg = ppg.cuda()
# 进行预测并数据转换到CPU
mel_pred, spec_pred = model(ppg)
mel_pred = mel_pred[0].cpu().data.numpy()
spec_pred = spec_pred[0].cpu().data.numpy()
# vocoder合成音频波形文件
mel_pred_audio = normalized_db_mel2wav(mel_pred)
spec_pred_audio = normalized_db_spec2wav(spec_pred)
return mel_pred, spec_pred, mel_pred_audio, spec_pred_audio