Grad-TTS的核心思想：把diffusion当做一个postnet（或者Plug-In）用于特征增强。因此，它可以是一种通用模块应用于任何网络中，典型的作为FastSpeech2的后处理模块。这里，作者以VITS的SVC场景为例，提供Grad-TTS融合进VITS的思想和代码。
思想：
1，训练原始VITS模型
具体实现，略~~~
2，训练插件Diffusion模型
1）冻结原始VITS模型所有参数
2）训练Diffusion模型学习Flow推理结果与wave后验编码结果Z之间的noise
3，Diffusion可以减少Flow推理结果与真值之间的Gap，可以减缓过平滑问题
代码：
VITS歌声转换中实现Plug-In-Diffsuion的代码（语音合成同样适用）：
https://github.com/PlayVoice/so-vits-svc-5.0/tree/plug-in-diffusion
满足MIT协议下，该代码的使用无限制
下面是架构原理图，与操作步骤
Plug-in diffusion based on Grad-TTS from HUAWEI Noah's Ark Lab
Base framework ~~~
Plug-In-Diffusion
Notices
It looks like it's useless, but it seems to be somewhat useful
好像没啥用，好像有点用
训练
-
Complete the training of the bigvgan-mix-v2 master model
完成 bigvgan-mix-v2 主模型的训练
-
Create a working path and pull the branch codes: different from the bigvgan-mix-v2
创建工作路径，拉取分支代码：与 bigvgan-mix-v2 不同
-
install additional dependencies for diffusion:
为 diffusion 安装额外依赖:
pip install einops
-
Copy bigvgan-mix-v2 training data data_svc and files to the current working directory: same as bigvgan-mix-v2 training data
拷贝 bigvgan-mix-v2 的训练数据 data_svc 与 files 到当前工作目录：与 bigvgan-mix-v2 训练数据一样
-
Specify the master model path in configs/base.yaml:
在 configs/base.yaml 中指定主模型路径:
pretrain: "bigvgan-mix-v2/chkpt/sovits5.0/sovits5.0_0500.pt"
-
Start train
启动训练
python svc_trainer.py --config configs/base.yaml --name plug
Check the log to be sure: your master model is loaded
python svc_trainer.py --config configs/base.yaml --name plug
Batch size per GPU : 8
----------10----------
2023-09-06 06:31:23,136 - INFO - Start from 32k pretrain model: sovits5.0_1100. pt
plug.estimator.spk_mlp.0.weight is not in the checkpoint
plug.estimator.spk_mlp.0.bias is not in the checkpoint
plug.estimator.spk_mlp.2.weight is not in the checkpoint
plug.estimator.spk_mlp.2.bias is not in the checkpoint
plug.estimator.mlp.0.weight is not in the checkpoint
plug.estimator.mlp.0.bias is not in the checkpoint
plug.estimator.mlp.2.weight is not in the checkpoint
plug.estimator.mlp.2.bias is not in the checkpoint
plug.estimator.downs.0.0.mlp.1.weight is not in the checkpoint
plug.estimator.downs.0.0.mlp.1.bias is not in the checkpoint
plug.estimator.downs.0.0.block1.block.0.weight is not in the checkpoint
plug.estimator.downs.0.0.block1.block.0.bias is not in the checkpointInference
python svc_inference.py --config configs/base.yaml --model chkpt/plug/plug_***.pt --spk ./data_svc/singer/your_singer.spk.npy --wave test.wavsvc_inference.py has a small changes from bigvgan-mix-v2