README.md

---
license: openrail
task_categories:
- conversational
language:
- aa
tags:
- music
size_categories:
- n<1K
pretty_name: genshin_voice_sovits 
---
# 效果预览

本仓库用于预览训练出的各种语音模型的效果，点击角色名自动跳转对应训练参数。</br>
正常说话的音色转换较为准确，歌曲包含较广的音域且bgm和声等难以去除干净，效果有所折扣。</br>

|  角色名  |  角色原声A   |     被转换人声B   |A音色替换B|A音色翻唱(点击直接下载)| 
| :------: | :----: | :----: | :----: |:----: | 
| [散兵](https://huggingface.co/datasets/jiaheillu/audio_preview/blob/main/散兵效果预览/训练参数速览.md)| <audio src="https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/散兵效果预览/部分训练集/真遗憾，小吉祥草王让他消除了那么多的切片，剥夺了我将他一片一片千刀万剐的快乐%E3%80%82.mp3" controls="controls"> | <audio src="https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/散兵效果预览/原声/shenli3.wav" controls="controls"> | <audio src="https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/散兵效果预览/转换结果/shenli3mp3_auto_liulangzhe.wav" controls="controls">|[夢で会えたら](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/散兵效果预览/转换结果/夢で逢えたら2liulangzhe_f.wav)|
|[胡桃](https://huggingface.co/datasets/jiaheillu/audio_preview/blob/main/胡桃_preview/README.md)| <audio style="https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/胡桃_preview/hutao.wav" controls="controls"> | .........| .........  |[moonlight shadow](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/胡桃_preview/moonlight_shadow2胡桃.WAV),[云烟成雨](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/胡桃_preview/云烟成雨2胡桃.WAV),[原点](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/胡桃_preview/原点2胡桃.WAV),[夢で逢えたら](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/胡桃_preview/夢だ会えたら2胡桃.WAV),[贝加尔湖畔](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/胡桃_preview/贝加尔湖畔2胡桃.WAV)  |
|[神里绫华](https://huggingface.co/datasets/jiaheillu/audio_preview/blob/main/绫华_preview/README.md)| <audio src="https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/绫华_preview/linghua428.wav" controls="controls">  | <audio src="https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/绫华_preview/yelan.wav" controls="controls"> | <audio src="https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/绫华_preview/yelan.wav_auto_linghua_0.5.flac" controls="controls">  |[アムリタ](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/绫华_preview/アムリタ2绫华.WAV),[大鱼](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/绫华_preview/大鱼2绫华.WAV),[遊園施設](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/绫华_preview/遊園施設2绫华.WAV),[the day you want away](https://huggingface.co/datasets/jiaheillu/audio_preview/resolve/main/绫华_preview/the_day_you_want_away2绫华.WAV)|
  

关键参数：

音频时长：min<br>
epoch: 轮 <br>

其余：
batch_size = 一个step训练的片段数<br>
segments = 音频被切分的片段<br>
step=segments*epoch/batch_size,即模型文件后面数字由来<br>

以散兵为例：
损失函数图像：主要看step 与 loss5，比如：<br>
给一个大致的参考，待转换音频都为高音女生，这是较为刁钻的测试：如图，10min纯净人声，<br>
差不多2800epoch（10000step）就已经出结果了，实际使用的是5571epoch(19500step)的文件，被训练音色和原音色相差几<br>
何，差不多有个概念。当然即使loss也不足以参考，唯一的衡量标准就是当事人的耳朵。当然，正常训练，10min还是有些少的。<br>

相关文件全部在“散兵效果预览”文件夹中<br>
![sanbing_loss](./散兵效果预览/%E8%AE%AD%E7%BB%83%E5%8F%82%E6%95%B0%E9%80%9F%E8%A7%88.assets/sanbing_loss.png)