MomoyamaSawa
commited on
Commit
•
07dbe8b
1
Parent(s):
ca1f1c1
update README
Browse files- GPT_weights/{nene100-1-e15.ckpt → nene60-1-e15.ckpt} +0 -0
- GPT_weights/{nene100-2-e15.ckpt → nene60-2-e15.ckpt} +0 -0
- GPT_weights/{nene100-test-e20.ckpt → nene60-test-e20.ckpt} +0 -0
- README.assets/stamp0570.png +0 -0
- README.md +32 -3
- SoVITS_weights/{nene100_1_e8_s640.pth → nene60_1_e8_s640.pth} +0 -0
- SoVITS_weights/{nene100_2_e2_s168.pth → nene60_2_e2_s168.pth} +0 -0
- SoVITS_weights/{nene100_2_e4_s336.pth → nene60_2_e4_s336.pth} +0 -0
- SoVITS_weights/{nene100_test_e8_s280.pth → nene60_test_e8_s280.pth} +0 -0
GPT_weights/{nene100-1-e15.ckpt → nene60-1-e15.ckpt}
RENAMED
File without changes
|
GPT_weights/{nene100-2-e15.ckpt → nene60-2-e15.ckpt}
RENAMED
File without changes
|
GPT_weights/{nene100-test-e20.ckpt → nene60-test-e20.ckpt}
RENAMED
File without changes
|
README.assets/stamp0570.png
ADDED
README.md
CHANGED
@@ -1,3 +1,32 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<p align = 'center'>
|
2 |
+
<img width='150' src='./README.assets/stamp0570.png'>
|
3 |
+
</p>
|
4 |
+
<p align = 'center'> 🥕 </p>
|
5 |
+
<p align = 'center'> 如果兔兔的仓库对你有帮助的话点个⭐喵~ </p>
|
6 |
+
<p align = 'center'> If Tutu's repository is helpful to you, please give it a ⭐ meow~ </p>
|
7 |
+
<p align = 'center'> もしうさぎのリポジトリが役に立った場合は、⭐をぽちっとしてくださいにゃん~ </p>
|
8 |
+
|
9 |
+
<p align = 'center'> 🍉 </p>
|
10 |
+
<p align = 'center'> 任何 ❓问题 / 💭思考 /💡想法 都欢迎提出!</p>
|
11 |
+
<p align = 'center'> Any ❓question / 💭thought /💡idea is welcome! </p>
|
12 |
+
<p align = 'center'> どんな ❓質問 / 💭考え /💡アイデア でも歓迎です! </p>
|
13 |
+
|
14 |
+
# 简介
|
15 |
+
|
16 |
+
基于 [GPT-SoVITS](https://github.com/RVC-Boss/GPT-SoVITS) 项目训练的 [草薙寧々](https://zh.moegirl.org.cn/%E8%8D%89%E8%96%99%E5%AE%81%E5%AE%81) 模型
|
17 |
+
|
18 |
+
经测试大部分情况下 nene30_e8_s328.pth > nene60_2_e4_s336.pth = nene60_test_e8_s280.pth > nene60_1_e8_s640.pth = nene60_2_e2_s168.pth,可以默认选择 **nene30_e8_s328.pth**,但在不同的情况下其他模型可能表现较好,在默认情况下表现不佳时可以切换模型 / 切换参考音频尝试
|
19 |
+
|
20 |
+
| SoVITS 模型 | 介绍 | 对应 GPT 模型 |
|
21 |
+
| :---------------------: | :-----------------: | :------------------: |
|
22 |
+
| nene30_e8_s328.pth | 30min 训练集 8epoch | nene30-e15.ckpt |
|
23 |
+
| nene60_1_e8_s640.pth | 60min 训练集 8epoch | nene60-1-e15.ckpt |
|
24 |
+
| nene60_2_e2_s168.pth | 60min 训练集 2epoch | nene60-2-e15.ckpt |
|
25 |
+
| nene60_2_e4_s336.pth | 60min 训练集 4epoch | nene60-2-e15.ckpt |
|
26 |
+
| nene60_test_e8_s280.pth | 60min 训练集 8epoch | nene60-test-e20.ckpt |
|
27 |
+
|
28 |
+
# 训练 & 推理
|
29 |
+
|
30 |
+
* 推理使用的参考音频对情感语气语调还有停顿影响很大,且必须使用原角色,不然效果会有点迷,建议不同情感语气语调抑扬,各种场合的干声存一份用作参考,平时大部分情况就用比较标准的陈述句当参考音频就好
|
31 |
+
* 根据项目作者得知训练集边际效应大概为 1h,经实验保持默认参数情况下 30min 的干声素材 8epoch 这样的训练量对于这个数据集大概是效果比较好的了,多的训练量比较容易过拟合,这边只做了随机 10min | 30min | 60min 数据量的实验,有时间也可以再做做更多时间和 epoch 的组合 / 更多次随机实验来试试有没有效果更好的
|
32 |
+
|
SoVITS_weights/{nene100_1_e8_s640.pth → nene60_1_e8_s640.pth}
RENAMED
File without changes
|
SoVITS_weights/{nene100_2_e2_s168.pth → nene60_2_e2_s168.pth}
RENAMED
File without changes
|
SoVITS_weights/{nene100_2_e4_s336.pth → nene60_2_e4_s336.pth}
RENAMED
File without changes
|
SoVITS_weights/{nene100_test_e8_s280.pth → nene60_test_e8_s280.pth}
RENAMED
File without changes
|