IndexTTS-2.5: 5-Language Zero-Shot TTS with Emotion Control

Overview

IndexTTS-2.5 is an autoregressive zero-shot TTS model built on a GPT backbone with a flow-matching speech-to-mel decoder and a BigVGAN vocoder (~0.8B parameters in the GPT backbone). It synthesizes 22.05 kHz speech from a reference voice prompt plus text, keeping the cloned timbre while controlling emotion independently.

Compared with IndexTTS-2, the new version adds Japanese, Spanish and Arabic (beyond Chinese and English), infers faster, adds speaking speed control, and improves controllability of pronunciation.

Key Features

• Zero-shot voice cloning — clone a voice from a single reference audio clip, no fine-tuning required.
• 5 languages — Chinese, English, Japanese, Spanish and Arabic, with cross-lingual voice transfer (speak any supported language in the cloned voice).
• Emotion control — an 8-float emotion vector [happy, angry, sad, afraid, disgusted, melancholic, surprised, calm], disentangled from timbre, adjusts the emotional delivery of the speech.
• Speaking speed control — adjust the pace of synthesized speech.
• Pronunciation control — Chinese Pinyin, English CMU phonemes, and Japanese Kana can be specified inline with <word|reading> notation for precise pronunciation.
• ~6 GB VRAM — runs on consumer GPUs with bf16 inference.

ComfyUI Support

IndexTTS-2.5 is not built into ComfyUI core, but community custom nodes are available:

• BSAI_ComfyUI_IndexTTS-2.5 — install into ComfyUI/custom_nodes/ to run IndexTTS-2.5 from the ComfyUI graph.
• ComfyUI_JR_IndexTTS25 — an alternative community node pack with its own WebUI.

Availability

The model is released on HuggingFace with an inference library and config, and the official IndexTTS-2.5 Demo space (Gradio) provides a no-code interface. The code repository also supports production deployment via vLLM recipes.

git clone https://github.com/index-tts/index-tts.git && cd index-tts
pip install -U uv && uv sync --all-extras
hf download IndexTeam/IndexTTS-2.5 --local-dir=checkpoints

Auxiliary models (w2v-bert-2.0, MaskGCT semantic codec, CAMPPlus, BigVGAN) are downloaded automatically on first run. The paper is available at arXiv:2601.03888.