Alibaba Unveils Qwen3.8-LiveTranslate — AI Translates Speech in Real Time and Tells Speakers Apart

The Qwen team at Alibaba has unveiled Qwen3.8-LiveTranslate, a new model for simultaneous real-time audio and video translation. The system supports 60 languages, can tell speakers apart, takes prior context into account, and voices the translation in a timbre close to that of the original speaker.

According to Qwen, average translation latency as measured by the LAAL metric dropped from 2.8 to 2.3 seconds compared with the previous generation. That does not mean every user is guaranteed a latency of exactly 2.3 seconds: the final time also depends on the connection, buffering and the playback method.

The model is built on a new architecture called Interleave. Instead of the classic “speech recognition → translation → voice synthesis” sequence, the system treats audio and text information as a single timeline. Audio that has already been processed and the translation it produced can be reused from context, which, according to the developers, both improves the coherence of the translation and reduces latency.

The AI tells multiple speakers apart and preserves their voices

One of the main new features is real-time speaker separation. When several people are talking, Qwen determines who each remark belongs to and keeps that attribution in the translation.

With voice output, the model also tries to reproduce each participant’s individual timbre. The feature is aimed primarily at meetings, interviews, conferences and other situations where it matters not only what is being said but who is saying it.

Qwen3.8-LiveTranslate can display the original transcript and the translation at the same time, and the long context helps the system handle names, specialized terms and ambiguous expressions whose meaning only becomes clear from earlier remarks.

The model also accepts images along with audio. QwenCloud’s documentation says visual information — for example, text on screen, gestures or other details of the video stream — can be used to refine the translation.

阿里刚刚放出了下一代实时同声传译模型：Qwen3.8-LiveTranslate，字均延迟从2.8秒压到2.3秒，支持60种语言

如果延迟稳定压到2.3秒这个水平，并且在多说话人、长音频场景下DER可控，那会议、展会、跨境电商谈判等大量这种“够用就行”的场景可能会率先被AI接管… https://t.co/Ih3oPNDO3H pic.twitter.com/3ztVv4yLuZ

— AIGCLINK (@aigclink) September 19, 2026

Voice translation is not available for all 60 languages

The stated 60 languages cover speech recognition and text translation. Synthesized translated speech is so far supported for 29 languages. The company lists the full set in the QwenCloud documentation.

You can try the technology through the LiveTranslate web demo. Developers can access the qwen3.8-livetranslate-flash-realtime model via API. Its context window is about 53,000 tokens, and international pricing starts at $7.50 per million input audio tokens; speech synthesis costs $30 per million output audio tokens.

In Alibaba’s own tests, the new version beat Qwen3.5-LiveTranslate and other systems on translation quality, speech recognition, latency and multi-speaker handling. For now these are Qwen’s own results, so the real gap with competitors has yet to be assessed in independent comparisons.