IndexTTS-2

IndexTTS-2 TTS

A zero-shot TTS model with fine-grained emotion control via emotion vectors, no emotion-specific training data required.

Sign up for 5,000 character limit

Wrap your text in SSML tags for precise control:

<speak><prosody rate="slow">Slow speech</prosody></speak>

Tags the selected model understands — click to drop one into your text where it happens:

This model reads plain text, so inline tags are ignored. For tag-based emotion, switch to an expressive model like Orpheus or Bark.

Define custom pronunciations (word = pronunciation):

-12 +12
0.5x 2.0x
Free with Piper, VITS, MeloTTS
Your generated audio will appear here. Choose a model, enter text, and click Generate.
Audio Generated Successfully
0:00
Download Audio Download .srt Link expires in 24h
Free tier: personal use. Commercial license from $5/mo
Love TTS.ai? Tell your friends!

About IndexTTS-2

IndexTTS-2, from the Index Team, is an expressive text-to-speech system that pairs zero-shot voice synthesis with precise emotional control. Rather than relying on emotion-labeled training data, it uses emotion vectors to dial in tones like happy, sad, angry, or fearful independently of the voice itself. Built on a Qwen2 backbone with BigVGAN as the vocoder, it supports English and Chinese and can clone a voice from roughly five seconds of reference audio. It suits audiobooks, virtual assistants, and any content where the same voice needs to shift emotional register. Its weights use the Bilibili Model License, which permits commercial use below large usage and revenue thresholds.

Best for: Emotionally expressive content, audiobooks, virtual assistants

Browse all IndexTTS-2 voices

At a glance

Developer
Index Team
License
Bilibili Model License
Tier
standard
Speed
medium
Voice cloning
Yes
Languages
English, Chinese
Max characters
1000

IndexTTS-2 voices

Chinese Default

Chinese
Standard Neutral

Default

English
Standard Neutral

IndexTTS-2 TTS — FAQ

It uses emotion vectors that let you specify tones such as happy, sad, angry, or fearful without needing emotion-specific training data, and the emotional expression is controlled independently from the voice identity.

Yes. It performs zero-shot voice cloning from a short reference, typically around five seconds of audio, in English or Chinese.

Its weights are released under the Bilibili Model License, which allows commercial use for products below defined user and revenue thresholds. Larger deployments should review the license terms.
← All voices