README.md

metadata

license: agpl-3.0
tags:
  - tts
  - text-to-speech
  - gpt-sovits
  - voice-clone
  - japanese
language:
  - ja

ATRI Voice Model — GPT-SoVITS v2Pro

WARNING: This model is for personal and research use only. Do not use it for commercial purposes or to impersonate real individuals.

---

Overview

A fine-tuned GPT-SoVITS v2Pro voice model for ATRI (from ATRI -My Dear Moments-), capable of synthesizing speech in Japanese, Chinese, and English.

Files

• ATR_e8_s3952.pth — Fine-tuned SoVITS model weights (8 epochs, 3952 steps)
• ref_audio.wav — Reference audio for inference
• api_atri.py — FastAPI-based TTS inference server

Usage

1. Clone and set up GPT-SoVITS following its instructions.
2. Download the GPT pretrained model s1v3.ckpt from GPT-SoVITS (included in its pretrained models).
3. Place ATR_e8_s3952.pth and ref_audio.wav in your preferred location.
4. Update the paths in api_atri.py (replace /path/to/ placeholders with actual paths).
5. Run the API server:

cd /path/to/GPT-SoVITS
python api_atri.py

API docs will be available at http://127.0.0.1:9880/docs.

API Endpoints

Endpoint	Method	Description
/health	GET	Health check
/tts	POST	Text-to-speech (returns full audio)
/tts/stream	POST	Streaming text-to-speech

Reference Audio

• Text: わたしはマスターの所有物ですので。 勝手に売買するのは違法です
• Language: Japanese

License

This project is licensed under AGPL-3.0, consistent with GPT-SoVITS.