tts

目前的多音字使用 pypinyin 或者 g2pM，精度有限，想做一个基于 BERT (或者 ERNIE) 多音字预测模型，简单来说就是假设某语言有 100 个多音字，每个多音字最多有 3 个发音，那么可以在 BERT 后面接 100 个 3 分类器（简单的 fc 层即可），在预测时，找到对应的分类器进行分类即可。
参考论文：
tencent_polyphone.pdf

数据可以用 https://github.com/kakaobrain/g2pM 提供的数据

进阶：多任务的 BERT
![image](https://user-images.githubusercontent.com/24568452

I'd like to train this model on 8 V100 GPUs - does it support multi GPU training?

since, at least in Arch Linux, each voice has its own package, the user can install rhvoice without voices, and when switching the synthesis module to RHVoice in orca, for example, the user will be left without speech.

I thought about rearranging items in sentence (using drag & drop) in case of misclick or user fault.