Index TTS 2.5

IndexTTS-2.5 is a zero-shot text-to-speech model from IndexTeam at Bilibili. It clones a voice from a single reference audio clip and separates emotion control from speaker timbre. The architecture is autoregressive: a GPT backbone of about 0.8B parameters, a flow-matching speech-to-mel decoder and a BigVGAN vocoder, producing 22.05 kHz audio.

It supports Chinese, English, Japanese, Spanish and Arabic, including cross-lingual voice transfer, for example dubbing Chinese speech into Spanish. Compared with IndexTTS-2, the model card lists added languages, faster inference, speaking-speed control (duration factor 0.5–2.0), and better control of pronunciation through Chinese Pinyin, English CMU phonemes and Japanese Kana. Emotion can be set with an 8-value vector, or from a text description via the QwenEmotion model.

The weights are downloadable from Hugging Face and ModelScope and run locally; inference requires Python 3.10–3.11, an NVIDIA GPU and roughly 6 GB of VRAM. A web UI is included. Limitations noted by the authors: long text is segmented and joined with short silences, so prosody is not modelled across segments. Random emotion sampling lowers cloning fidelity. The model does not verify speaker consent, which is left to the user. Use is governed by the bilibili Model Use License Agreement rather than a standard open-source license.