IndexTTS

Bilibili's Index team open-sourced IndexTTS in March 2025 and has shipped steadily since, through 1.5 in May 2025, IndexTTS-2 in September 2025 and IndexTTS-2.5 on 10 August 2026. The current model is 0.8B parameters covering Chinese, English, Japanese, Spanish and Arabic. Start with the licence, because it is what catches people out. The LICENSE file is Apache 2.0 text with an added bilibili clause: commercial use requires registration and prior written authorization, and you may not use the model or its derivatives to improve any other AI model except IndexTTS itself, its derivatives or non-commercial models. That is source-available, not open source, and there is an open issue on the repository from users asking the maintainers to say so in plain terms.

What you get for accepting the constraint is emotion control nothing else in the open field matches. Five separate paths are available: supply a second reference clip carrying the emotion you want, dial intensity with emo_alpha between 0.0 and 1.0, pass an 8-float vector specifying each emotion's weight, let the model infer emotion from the text itself or write an explicit emotion description. Duration is controlled independently through duration_factor across a 0.5x to 2.0x range, which is the feature that makes it viable for dubbing to a fixed timecode. Pronunciation can be forced with Pinyin for Chinese, CMU phonemes for English and Kana for Japanese, which handles the polyphone problem that wrecks most Chinese TTS on personal and place names.

Measured zero-shot performance averages 6.75 WER against 73.18 speaker similarity, breaking down to 4.36 WER with 77.10 SS on Chinese and 5.12 WER with 68.06 SS on English. Cross-lingual, using a Chinese reference to read English, gives 3.62 WER. Inference runs at 0.2065 real-time factor on an RTX 4090 in bf16, with fp16 and bf16 half precision, optional DeepSpeed acceleration and a vLLM path for production serving. Weights are on Hugging Face and ModelScope, so acquisition is easy from anywhere. The hard part is the commercial route, which means emailing [email protected] and negotiating in Mandarin. For a Chinese studio that is a phone call. For a team in Lagos or London it is a genuine obstacle, and it is the reason to reach for Fun-CosyVoice first unless you specifically need the emotion controls.