deepspeech

Specs

• Leon version: latest
• OS (or browser) version: Fedora 30
• Node.js version: 10.16.3
• Complete "npm run check" output:

➡ Here is the diagnosis about your current setup
✔ Run
✔ Run modules
✔ Reply you by texting
❗ Amazon Polly text-to-speech
❗ Google Cloud text-to-speech
❗ Watson text-to-speech
❗ Offline text-to-speech
❗ Google Cloud speech-to-text
❗ Watson spee

As implemented in Python in

alphacep/vosk-api@5e46825

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

We could include new corpora, works and texts in MITADS Dataset to increase the size of its vocabulary.

In 2021, all works written by people who died during the 1950s are released from copyright.

We could be collected works of these Italian writers died in 1950:

• Giovanni Paneroni, Italian writer
• Gaetano Pitta, Italian writer and journalist
• Tullio Giordana, Italian write