Alibaba drops an open-source voice cloner with 97ms latency

Qwen3-TTS is Alibaba's latest open-source text-to-speech model, engineered for high-speed voice generation and nuanced vocal control. Give it just a few seconds of reference audio, and it clones any voice on the fly. If you don't have audio on hand, you can describe a persona from scratch with a text prompt like "a raspy old man" or "a cartoon chipmunk."

What it can do

Alibaba rolled out two sizes: a full-featured 1.7B model that delivers top performance and full control, and a lighter 0.6B version optimized for balanced speed and efficiency.

Under the hood, a dual-track architecture yields the first audio chunk the moment a single character is entered, shaving end-to-end latency down to just 97 ms. That low threshold turns it into a serious contender for live-streaming bots, real-time speech translation, and interactive AI customer support.

Its cross-lingual range is equally versatile. Clone a Chinese voice sample, and it can speak English, Japanese, Korean, German, French, Russian, Spanish, Portuguese, Italian, or regional Chinese dialects like Sichuanese and Beijing dialect while keeping the original timbre intact, complete with emotional cues like sadness, joy, and anger.

Why it matters

For builders creating audiobooks, digital avatars, or real-time voice agents, a sub-100ms open-source model makes natural two-way conversation practical without paying steep fees for proprietary APIs.

Packing prompt-based voice creation, emotion control, and multi-dialect cloning into models under 2B parameters is a remarkably practical move. If your workflow needs low-latency voice, this is definitely worth spinning up locally.