PaddlePaddle/PaddleSpeech

[TTS] Try to train a universial GAN Vocoder using CSMSC + LJSpeech + AISHELL3 + VCTK

Open

#2,803 创建于 2023年1月6日

在 GitHub 查看
 (0 评论) (1 反应) (1 负责人)Python (9,453 star) (1,702 fork)batch import
T2Sfeature requestgood first issue

描述

An universial GAN Vocoder may works well for all AMs of different datasets, for example, CSMSC is a single female dataset, may generate bad wavs for mels of male speakers, cause different genders have different distribution of speech features.

Please try to train a universial GAN Vocoder using CSMSC + LJSpeech + AISHELL3 + VCTK + some other TTS datasets (if you want) with the config of CSMSC (24kHz).

LJSpeech is 22.05kHz, but you don't need to resample it yourself, cause we will resample the wavs to the sample rate setted in config file in preprocess stage ~

贡献者指南

技术栈
python
领域
machine learningai
议题类型
research
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
4
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
over 1 week
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
blocked
清晰度议题是否清楚说明期望改动、验收标准和下一步。
clear
前置要求
Knowledge of TTSGAN vocoder conceptsPythonPaddleSpeech
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
20
研究方向
This issue asks to train a universal GAN vocoder using multiple datasets (CSMSC, LJSpeech, AISHELL3, VCTK) with the CSMSC config (24kHz). The assignee is likely working on this. To start, review the existing GAN vocoder implementation in PaddleSpeech, such as the config file under examples/csmsc/tts3/conf/default.yaml. Prepare to handle different sample rates (LJSpeech is 22.05kHz but preprocessing resamples). Combine datasets and modify the training pipeline to support multi dataset training. Check if any open PRs or branches exist for this task.