Shiki Natsume (Shiki Natsume) — GPT-SoVITS Voice Model
Model Overview
A Japanese TTS (Text-to-Speech) voice model based on the GPT-SoVITS v2 ProPlus architecture, designed to generate speech for the game character Shiki Natsume (Shiki Natsume).
Character Background
| Attribute | Details |
|---|---|
| Name | Shiki Natsume (四季ナツメ) |
| Voice Actor | Kanako (夏和小) |
| Source | Yuzusoft visual novel Café Stella and the Reapers' Butterflies (喫茶カフェステラと死神の蝶) |
| Language | Japanese |
Note: This model is intended for personal learning, research, and non-commercial use only. Do not use the model or its outputs for any commercial purposes or unauthorized distribution.
Model Files
This repository contains the following model weights and reference audio files:
| File Path | Description |
|---|---|
GPT_weights_v2ProPlus/natsume.ckpt | GPT model weights |
SoVITS_weights_v2ProPlus/natsume.pth | SoVITS model weights |
RefAudio/casual/ | Casual tone reference audio |
RefAudio/romantic/ | Romantic tone reference audio |
Reference Audio Transcription
Casual tone (casual.ogg):
「ワタシ一人だけじゃなく、周りのみんなに助けてもらいながら考えて……こういうお店にしました。」 (Translation: "Not just by myself, but with everyone around me helping me think... I decided to open this kind of shop.")
Romantic tone (romantic.ogg):
「ワタシだって、好きな人と一緒にいたいと思うし。こうして触れ合ってると……幸せで。」 (Translation: "I also want to be with the person I like. When we're like this, touching... I feel so happy.")
Usage
Training
Use GPT-SoVITS for model training. The project provides comprehensive documentation for the training process.
Inference
Use GPT-SoVITS_minimal_inference for voice synthesis. This lightweight inference tool provides a simple API for generating speech with the model weights.
For detailed usage instructions, refer to the respective project documentation linked above.
Training Details
Training Framework
- Architecture: GPT-SoVITS v2 ProPlus
- Training Code: GPT-SoVITS
Training Data
- Source: Extracted from the visual novel Café Stella and the Reapers' Butterflies
- Training Tool: Dataset Maker for Galgames
- Data Volume: Approximately 140 minutes of high-quality speech
- Format: Ogg Vorbis
Limitations and Risks
Technical Limitations
- Language: The model is optimized for Japanese only; performance on other languages may be poor
- Emotion Range: Currently supports only casual and romantic emotional styles
- Audio Length: Single inference is recommended to be under 30 seconds; longer text may result in quality degradation
Usage Risks
- Copyright: The model is trained on game voice data and should only be used for personal learning and non-commercial research
- Content Safety: Do not use the model to generate inappropriate content or for fraudulent purposes
- Voice Mimicry: The model-generated voice may resemble the original voice actor; please respect the rights of the original creators and voice actors
License
This Model
- License: Non-commercial use only (Other)
- Prohibited: Any commercial use, paid distribution, or training other commercial models
- Allowed: Personal learning, research, non-commercial sharing
Third-Party Components
This model relies on the following open-source projects:
| Component | License | Link |
|---|---|---|
| GPT-SoVITS | MIT | GitHub |
| GPT-SoVITS_minimal_inference | Apache License 2.0 | GitHub |