LJSpeech training result

[WendongGan]

@p0p4k hi , p0p4k, thanks again for your shared code.

I have trained the two configure in LJSpeech about 300k+ steps , with batch size 32, 3090Ti

LJ Speech

python train.py -c configs/vits2_ljs_nosdp.json -m ljs_base # no-sdp;
python train.py -c configs/vits2_ljs_base.json -m ljs_base # with sdp;

Result samples
VITS2_LJSpeech Result_from WendongGan-2023.08.29.zip

Pretrained model:
(coming soon.....)

Now, We think no-dsp is better, maybe it is more easy to train. When i finish my training, i will update the samples

[p0p4k]

Thanks for the samples and the update. Do you think it is better than VITS-1?

[WendongGan]

yes, I think it's better than VTS-1 in terms of emphasis and pauses. What do you think? @p0p4k

[p0p4k]

yes, I think it's better than VTS-1 in terms of emphasis and pauses. What do you think? @p0p4k

I think so too!

[harmlessman]

How is the synthesis speed compared to vits1? Is the vits2 much faster?

[EobardThawne721]

@p0p4k hi , p0p4k, thanks again for your shared code.

I have trained the two configure in LJSpeech about 300k+ steps , with batch size 32, 3090Ti

LJ Speech

python train.py -c configs/vits2_ljs_nosdp.json -m ljs_base # no-sdp; python train.py -c configs/vits2_ljs_base.json -m ljs_base # with sdp;

Result samples VITS2_LJSpeech Result_from WendongGan-2023.08.29.zip

Pretrained model: (coming soon.....)

Now, We think no-dsp is better, maybe it is more easy to train. When i finish my training, i will update the samples

Hello, why is it that I've run 450k steps on a single GPU using the code provided, only changing the batch size to 8, but still see no improvement?（你好，为什么我用他提供的代码，我只更改了batch为8，在单个gpu上运行了450k步还是没效果呢）

[JohnHerry]

You should better adjust your batch size. the value of 8 is too small. and, multi-card training helps the smooth traning。

[EobardThawne721]

Can it only run on multiple GPUs? Can't it run on a single GPU? What is the usual batch size for multi-GPU training？