Back to Gap Map Model components / Speech & audio

IndexTTS

bilibili
open weights / Overall score: 2.6

Bilibili's zero-shot text-to-speech model, which clones a voice from a single reference clip. IndexTTS-2.5 speaks Chinese, English, Japanese, Spanish and Arabic, carries a voice across languages, controls emotion separately from timbre, sets speaking speed, and accepts pinyin, CMU phoneme or kana hints for pronunciation.

Openness

3 high confidence
3.0
weights
open (IndexTTS-2.5 checkpoints on the Hub, ungated)
data
closed (no training-data statement on the card or the README)
code
partial (inference code and a web UI; no training pipeline)
license
bilibili-Model-Use-License (IndexTTS-2.5 weights and the repository code; a separate license is required above 100 million monthly active users or RMB 1 billion annual revenue)
superseded-release
IndexTTS-2 (2025, the model the bilibili license was written for; IndexTTS-1.5 before it is tagged Apache-2.0 on the Hub)

IndexTTS-2.5 downloads freely, and the bilibili Model Use License allows commercial use for everyone except the very largest companies, which must ask for a separate license. Bilibili publishes inference code only and says nothing about the training data.

Adoption

2 high confidence
2.0

Hugging Face downloads over the trailing month, summed across the IndexTTS-2.5, IndexTTS-2 and IndexTTS-1.5 checkpoints. The ModelScope copies are not counted.

Capability

3 low confidence
3.0

IndexTTS has no listener rating yet, so it is placed on its own results on the public CV3-Eval test set, where its figures sit close to CosyVoice 3, OmniVoice and VoxCPM2.

Verified 2026-11-12