IndexTTS 2.5 drops with 5-language zero-shot cloning

IndexTTS-2.5 is an open-source zero-shot text-to-speech system designed to clone any voice from a short reference audio clip. Released on August 10, 2026, the updated model adds support for Chinese, English, Japanese, Spanish, and Arabic while delivering faster inference and cleaner separation between timbre and emotion.

What you can do with it

The core workflow is dead simple: feed it a few seconds of voice audio, and the model can speak whatever script you give it. This makes cross-lingual dubbing painless, allowing a creator to take a native Chinese recording and make it speak natural Spanish or English in the exact same voice.

Controllability takes center stage in this release. You can adjust speaking speed through a duration factor between 0.5x and 2.0x without pitch distortion, giving editors precise pacing control.

Pronunciation accuracy also gets a major upgrade. To handle tricky polyphones and regional accents, IndexTTS-2.5 lets you inject Chinese Pinyin, English CMU phonemes, and Japanese Kana directly into the script using bracketed annotations.

Emotion handling is equally versatile. You can guide emotion using a separate audio reference with an adjustable weight slider, manually dial in an eight-dimensional emotion vector covering states like anger and joy, or enable automatic emotion inference from the raw text powered by Qwen.

How to get it running

The code is live on GitHub, with model checkpoints hosted on both Hugging Face and ModelScope. The full weight package sits around 5.5 GB, meaning you can easily run it on everyday consumer GPUs.

Setup is streamlined using the uv package manager, which provisions an isolated environment with dependencies automatically. A built-in Gradio WebUI is included, accessible at localhost port 7860 right after launch.

For inference efficiency, IndexTTS-2.5 supports BF16 half-precision out of the box to keep VRAM usage lean. Linux users can enable DeepSpeed acceleration or compiled CUDA kernels for lower latency, while enterprise deployments can leverage the official vLLM recipe for production-grade throughput.

Why it matters

For solo creators, indie game developers, and localization teams, IndexTTS-2.5 is one of the most flexible open-source voice engines available today. It packs multi-language voice cloning, phoneme-level pronunciation steering, and nuanced emotional range into a lightweight footprint that requires zero fine-tuning.

What makes this release stand out is its production mindset. Instead of staying trapped inside academic benchmark scripts, the project offers native vLLM serving and half-precision support, solving the real-world latency and concurrency bottlenecks that usually keep open-source speech models out of production.