SteinsGateSg/mayuri-voice

Mayuri Voice Model

Final Hugging Face packaging for the Shiina Mayuri GPT-SoVITS voice model.

Contents

• models/gpt/mayuri_v2-e8.ckpt Final GPT checkpoint selected after comparison.
• models/sovits/mayuri_v2_e20.pth Final SoVITS checkpoint selected after comparison.
• configs/ Training configs used to produce the final pair.
• refs/ Curated reference bank grouped by emotion, with top-ranked reference clips and matching text files.
• metadata/final_assets.json Final asset manifest.
• metadata/mayuri_profile.yaml Lightweight profile for downstream tooling.

Recommended Final Pair

• GPT: models/gpt/mayuri_v2-e8.ckpt
• SoVITS: models/sovits/mayuri_v2_e20.pth

Reference Bank

Use refs/index.csv to find recommended reference clips by emotion:

• neutral
• gentle
• happy
• excited
• worried
• sad
• teasing
• serious
• embarrassed
• other

Each reference entry contains:

• a .wav clip
• a matching .txt transcription
• ranking metadata in refs/index.csv

Notes

• This repository packages the final inference assets only.
• Full training logs and intermediate checkpoints are intentionally excluded.
• The original working repository also contains dataset preparation scripts, training wrappers, and experiment docs.

Upload Notes

• This repository is prepared for Git LFS upload.
• Run git lfs install before pushing large files.
• Weights and reference audio are tracked via .gitattributes.

Chinese Note

这个目录是面向 Hugging Face 模型仓库的整理结果，已经把最终选定的 SoVITS e20 + GPT e8、推荐参考音频库、配置文件和 profile 一起整理好了，可以直接初始化成独立仓库并上传。