| # Checkpoints Layout | |
| All inference checkpoints should be placed under `checkpoints/`. | |
| ```text | |
| checkpoints/ | |
| stage1_knnvc/ | |
| hifigan/ | |
| g_00125000.pt # required for stage1 | |
| wavlm/ | |
| WavLM-Large.pt # required for stage1 | |
| stage2_sovits/ | |
| chkpt/ | |
| sovits5.0_all_nocross/*.pt # default non-cross inference | |
| sovits5.0_all_cross/*.pt # cross-mode path in code | |
| whisper_pretrain/ | |
| large-v2.pt | |
| hubert_pretrain/ | |
| hubert-soft-0d54a1f4.pt | |
| crepe/assets/ | |
| full.pth | |
| speaker_pretrain/ | |
| best_model.pth.tar | |
| vits_pretrain/ | |
| sovits5.0.pretrain.pth | |
| ``` | |
| Notes: | |
| 1. If your cross checkpoint directory is `sovits5.0_all_cross/`, pass it explicitly with `--model` (`stage2`) or `--stage2-model` (`pipeline`). | |
| 2. `speaker_pretrain/best_model.pth.tar` is used by `build-spk`; it is not required when you already have ready-made `.spk.npy` files. | |
| For attribution and citation notes, see: | |
| - `THIRD_PARTY_NOTICES.md` | |
| - `docs/model_attribution.md` | |