search

keonlee9420 / PortaSpeech

PyTorch Implementation of PortaSpeech: Portable and High-Quality Generative Text-to-Speech
MIT License
331 stars 36 forks source link

无法训练 #7

Closed aijianiula0601 closed 2 years ago

aijianiula0601 commented 3 years ago

请问下，现在的程序是不是训练不了呢？ 在text的处理上，代码采用的是字符作为单元，而不是空格分开的音素为单元。发现跟phones_per_word对应不上。就是 len(phones)!=sum(phones_per_word)，我分析到的原因是phones采用的字符为单元。

keonlee9420 commented 3 years ago

Hi @aijianiula0601 , the current implementation doesn't support other languages and datasets than LJS yet; hence you might have to apply the proper boundary function for them. I may misunderstand your question since I'm answering by running Google translator. I can answer you better in English.

keonlee9420 commented 2 years ago

Close due to inactivity.

PiotrowskiD commented 2 years ago

Hi, I have the same problem is there a chance you could help me fixing this? Our dataset is english too and I've also used mfa for alignment. I did however change some of the config fields to adjust - sampling rate, filter length, hop length and window length. I've also used the same lexicon (so I guess maybe the problem comes from words present in our dataset and not in the lexicon?)