Cheatsheet for training voice models using so-vits-svc-fork

https://github.com/voicepaw/so-vits-svc-fork

1. Make a folder with the name of the voice you're training, cd to that directory
2. Put voice samples in a subdirectory named dataset_raw_raw
3. Run svc pre-split, this will generate dataset_raw
4. Run post_split.sh, this will split all samples from dataset_raw into clips of 30 second max length, and put them in dataset/44k/$(pwd)
5. Run svc pre-config, this will generate configs/44k/config.json
6. Edit configs/44k/config.json, modify epochs to some reasonable value like 500. Modify keep_ckpts to be the number of backlogs you want. I usually pick 5.
7. Run svc pre-hubert, this will generate stuff in dataset. Note that clips were split to max 30 second length because hubert takes more VRAM the longer the clip is. If you run out of VRAM you'll have to adjust the max clip length in post_split.sh
8. Run svc train, this will generate the models in logs\44k. Use the G_ models and config.json for inferrence.