Author / Maintainer: Bilibili @十一君Jun

Video demos and update news are all on Bilibili. You can also leave questions and feedback there.

A locally adapted version of the official IndexTTS2 model. It replaces the official gradio WebUI with a local workbench built on FastAPI + vanilla JavaScript: the service starts instantly, idle time doesn’t occupy VRAM, generation shows five-step real-time progress, results include waveform playback and a “crop / volume / speed” editing panel, and segment-level SRT / VTT subtitles can be exported. Presets and terminology pronunciations can be saved and reused.

Measured on RTX 3060 Ti 8GB: for the same test text, total generation time is reduced to about 1/9 of upstream (normalized to 12.77 seconds of audio of the same length: 151 s → about 17 s, RTF 11.87 → 1.31). For a Chinese long-form text of about 500 characters, with 25 diffusion steps / CFG 0.7 / fp16 all enabled: