An HTTP API that wrangles VITS, Bert-VITS2, and GPT-SoVITS

What it does

vits-simple-api is a Python HTTP API that wraps several popular neural text-to-speech and voice-conversion models—VITS, Bert-VITS2, GPT-SoVITS, HuBert-soft, and emotional variants—behind a single set of endpoints. It handles model loading, automatic language detection, long-text batching, and GPU inference so you don’t have to wire up each research codebase yourself. A browser-based inference UI and an admin backend for swapping models are included out of the box.

The interesting bit

The project started as an extension of Moegoe and has evolved into integration glue: it abstracts away the idiosyncratic configs and data formats of each underlying engine, letting you hot-swap voices or mix narrators and quoted speakers via URL parameters. That makes it particularly handy for reading apps like Legado that expect a simple, predictable audio endpoint.

Key highlights

• Supports multiple TTS engines simultaneously (VITS, Bert-VITS2, GPT-SoVITS, W2V2/emotional-vits, and Chinese VITS variants).
• Auto-detects language and cleans text according to each model’s expected input format.
• Includes a WebUI for generation and an admin panel for loading/unloading models without restarting.
• Offers GPU acceleration and batch processing for long inputs.
• Provides a reading API with separate narrator and in-quote speaker parameters.

Caveats

• SSML support is marked work-in-progress.
• The author notes they haven’t tested the Linux GPU installation steps.
• Documentation is functional but fragmented; some sections read like raw notes rather than a finished manual.

Verdict

Worth a look if you need a self-hosted, multilingual TTS server for reading apps or small creative tools. Skip it if you want a polished SaaS or a single-model research playground.