Wraps VITS, Bert-VITS2, GPT-SoVITS, HuBert-soft VITS and W2V2 emotional VITS models behind an HTTP API (GET/POST under /voice/) and a browser inference page. It loads multiple models from data/models, detects text language automatically, and supports voice conversion, long-text batching, streaming and optional GPU inference. An admin panel handles loading and unloading models.
Who it is for: Self-hosters who need a TTS API for VITS or GPT-SoVITS voice models
Strengths
Serves several VITS variants, including Bert-VITS2 and GPT-SoVITS, from one API
Output formats include wav, ogg, silk, mp3 and flac, with a streaming option
Auto-loads every model folder in data/models; admin panel for load and unload
Optional API key auth; admin backend can be disabled in config.yaml
Weaknesses
No models bundled; you download and place VITS models yourself
Bert-VITS2 models need a version field in config.json for compatibility
Linux GPU setup is untested by the author, per the README
Last release is 2025-02-03; AGPL-3.0 license limits closed-source network use
What it needs
GPU optional
Docker + Compose
Needs Python 3.10, PyTorch, Docker (optional), CUDA (for GPU)