About ZS-TTS dataset selects

[Lizh-TW]

Hi According to my recent experiments because (figure 0), I have something experience about select the dataset on ZS-TTS.

• figure 0
  

• First, I think the speaker have wav files count and the speaker's TTS similarity it not have a direct relationship(see figure 1)

  • figure 1

• Second, base the first, so I try to add Mozilla dataset in my training data, Then I got good result in unseendata. (See figure 2 data is my classmate. figure3 is the official YourTTS and use data is VoxCeleb1)

  • figure 2
  • figure 3

• Finally, I found that different accents need to be trained separately, I have tried training in Chinese (China and TW)

Thanks for reading and hope to get a reply 😀

[erogol]

Thanks for sharing your results. Really interesting.

One question. What did you use to measure speaker similarity?

[Lizh-TW]

@118dz 你有嘗試過使用官方接學訓練一種單speaker模型嗎?先從這部分開始吧

[118dz]

就是按照官方的那个教程下了训练集，但是吧感觉那个好像没跟新了，有些方法和变量它不适用了，然后就是一直报错

[Lizh-TW]

@118dz 抱歉那我可能無法幫助你，我已經近1年沒有碰此項目了，身邊也沒有足夠效能的電腦來測試。也許你可以把抱錯貼出來詢問作者

[118dz]

那您当初训练出来的数据集还在吗？训练出来的数据集怎么加载到程序中，有没有当初的一些笔记啥的，真心想学习一下，

[118dz]

哈哈，可算是把他那个demo跑起来了，环境有点问题高了，降低了版本就行了，参数给的不对也改了，但是还没开始训练中文，下载的训练集csv和demo中的csv感觉结构有点不一样，这个要怎么修改是该代码还是怎么怎么操作？想请教一下，还有就是训练结束后怎么使用该模型，谢谢大佬的回复