babysor / babysor/MockingBird

第二次贡献模型,同时谈谈相关经验。

Open
#245 13 comments 14 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
36.9k
Forks
5.2k
PR merge metrics
No merged PRs in 30d

Description

先上图再说
![attention_step_70000_sample_1](https://user-images.githubusercontent.com/22911884/144376652-d4a0c96e-1f48-4387-a978-b9d141bf789b.png)
![step-70000-mel-spectrogram_sample_1](https://user-images.githubusercontent.com/22911884/144376663-b92c2163-f9e9-4cc4-b593-b59860878cb4.png)

**aidatatang_200zh以及aishell3两个数据集,混合训练,batch size 96。**
在训练40K就可以达到0.18了,不过我觉得还是多训练一会好, **所以分享的这个是70K、0.2的模型。**

链接:https://pan.baidu.com/s/17yWmyq6_rh5MbCOwE3hH2Q
提取码:7777
- [ ] 最新版本可用

接下来为个人训练模型经验,可不看。

- [ ] 1、batch size对于loss值影响巨大,同样的混合训练,12和96这两个值会得到完全不同的结果,体现在:默认batch size 12的情况下,loss值很难达到0.35以下,而且不停的波动,且波动数值巨大。比如0.3跳0.5再跳回0.3。故个人建议尽可能开到比较大的batch size。32G显存在训练混合数据集时,最高稳定值在96,供参考。

- [ ] 2、本项目对游戏角色语音,例如王者荣耀语音克隆,无法得到好的结果。体现在模型克隆出来出现严重的机器音,或者音色根本不像。个人猜测是因为提供的4个数据集本身就是日常的一个对话,并没有游戏角色语音的情感等,所以无法很好克隆游戏的角色语音。

- [ ] 以上仅为萌新的个人经验,不代表是对的,也很有可能是错误的,仅供参考。

Contributor guide

No contributing guide indexed for this repository

Research direction

The issue provides a Baidu model download and notes about mixed-dataset training, batch size, and voice cloning, but names no source files or tests. Start by reviewing the shared model and training notes; the issue does not define an implementation task or a clear completion condition.

Written by the indexing model from the issue text.

Assessment

Domain
machine-learning
Issue type
Documentation
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
15/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.