aijianiula0601
aijianiula0601
I have a question. The input_lengths has not send to calculate the mask for mulithead-attention. Is it work?
So am I. It's great job. my email: [email protected]
How to build the predict graph? Trouble to understand loading the trained viarbles to decode,because the decode variables in a for loop.Thanks!
why the variable named as "pathBuuffer" already has 2 elements? actually is empty in my mind.
> > > 中文的你得拿中文数据集重新训练，亲测有效~ > > > > > > 您好，我用中文数据集训练了，效果也是不好，能麻烦宁分享下训练技巧么？ > > @zhangsong427 > 您好，我用的是AISHELL1数据集训练，用griffin lim得到得音频质量确实很差，所以我自己结合了WaveRNN作为vocoder来由Mel频谱合成语音，音质和音色转换效果还挺不错。 > > 但是我个人认为One-shot+WaveRNN也只能达到demo的效果，无法大批量稳定，因为我发现One-Shot没有利用语音的f0信息，导致对于某些语音，音色转换后的语音有些字会跑调~ 您好，用AIshell1来训练的话，有测试过不在训练集中的其他数据源比如THCHS-30的音色转换效果吗？AIshell1的数据集只有300来个speaker，训练出来的Speaekr encoder真的能找到unseen speaker的分布吗？谢谢。
The code read hardly.The organization of code make me great effort to read.
I have the same problem.Had it solved?
I got the same error