PlayVoice / lora-svc

singing voice change based on whisper, and lora for singing voice clone
MIT License
630 stars 78 forks source link

Diff-svc, so-vits-svc效果对比 #58

Open yygg678 opened 1 year ago

yygg678 commented 1 year ago
  1. lora-svc 的效果有跟Diff-svc, so-vits-svc效果对比吗?我看他们用的soft-vc 特征, 多说话人svc 时音色有泄漏,whisper 的ppg 多人有这个问题吗?
  2. lora-svc 是否都支持转语音转换和唱歌转换,无论训练的speaker 数据是只有歌唱还是只有语音数据?
MaxMax2016 commented 1 year ago

1,音色泄漏应该没有吧,可以用发布的模型进行测试,模型使用56个发音人训练 2,不支持语音转换,但是支持用语音数据训练歌声转换

yygg678 commented 1 year ago

用发布的模型进行测试,singer0001 唱高音时破音了,音高pitch是1000多,针对这种情况,有什么好的方法吗?

MaxMax2016 commented 1 year ago

https://github.com/PlayVoice/lora-svc/blob/main/svc_preprocess_f0.py#L13 模型训练的时候,设置了最高音为900,需要根据您的实际数据修改这个参数来训练模型

KillauraHacks commented 1 year ago

https://github.com/PlayVoice/lora-svc/blob/main/svc_preprocess_f0.py#L13 During model training, the highest pitch is set to 900. You need to modify this parameter according to your actual data to train the model

how do you find the highest pitch in your model and how do you find the number to use (eg. 900) that the pitch corresponds to in order to modify the parameter

Liujingxiu23 commented 1 year ago

@yygg678 请问你实际对比操作了不,该方法和Diff-svc, so-vits-svc对比效果如何?

yygg678 commented 1 year ago

@yygg678 请问你实际对比操作了不,该方法和Diff-svc, so-vits-svc对比效果如何? 看数据集吧,我数据集比较脏,so-vits-svc 好些,diff-svc 没有试试