IndexTTS

Overview

IndexTTS family of models can copy a voice from a single clip. You give IndexTTS a sample and it’ll make that voice say anything you want. No extra training. Just drop in a reference and go.

Want the speaker to sound excited, or mad, or chill? You can feed it another clip with the right vibe or just tell it in your text prompt. There’s also a little emo-alpha knob to dial up how strong the emotion should be.

With version 2 you can tell it exactly how long the line should be or let it run wild. Handy for syncing lips or matching scenes. Just know this super precise mode still isn’t fully ready yet according to the devs.

Supports more than one language and can handle mixed stuff like Chinese characters and Pinyin in one go. That means pronunciation control gets way easier if you’re working with Mandarin.

Supports FP16 faster computing with DeepSpeed and CUDA kernels for speedups. But you’ll need a GPU and proper setup. CUDA 12.8+ is a must if you want full power.

So what about the licence terms? Let's take IndexTTS-2 specifically.

Commercial use is allowed for individuals and smaller/medium businesses.

Unrestricted use is not allowed for very large platforms (100M+ MAUs or >RMB 1B revenue) unless you negotiate a separate license.

You can’t use it to boost/train other commercial AI models beyond indextts2 or its derivatives.

Supported Languages

• Chinese
• English