Alibaba’s Qwen team released a real-time interpretation model on September 18 built to fix two separate problems at once: how fast a system translates speech, and whether it can tell who is talking. The company says average lag fell from 2.8 seconds in the prior generation to 2.3 seconds, while the new model, Qwen3.8-LiveTranslate, gained the ability to track individual speakers across a conversation and keep each one’s voice consistent in the translated output.
That speed gain rests on a rebuilt architecture the team calls “Interleave.” It treats what a listener hears and what the system has already translated as one continuous audio-text stream, so earlier audio and earlier translations can be cached and reused instead of reprocessed from scratch. Qwen’s engineers credit this design for the faithfulness and fluency gains that came alongside the latency drop, though the company has not published the numbers behind those quality claims.
Three capabilities sit on top of the speed work. The model separates speakers in multi-person conversations, attributing each sentence to the right person and keeping that person’s cloned voice steady across the translated audio. It also puts the original and translated text side by side on one screen, a format meant to let a listener check the translation against the source as it happens. And it draws on earlier turns in the conversation, plus visual context, to resolve names and homophones that a sentence-by-sentence translator would get wrong.
Under the hood, Qwen built the system as a Hybrid Mixture-of-Experts design split into two parts. A “Thinker” component sequences text, audio, translation and video into one time-ordered stream, and a “Talker” component turns that sequence into spoken output, delivered in the original speaker’s own timbre. Input coverage runs to 60 languages; the spoken output covers 29.
Qwen’s own tests are the only evidence offered for how the model stacks up against rivals. On a multi-speaker set the company calls Omnilingua-MSpeaker, spanning 14 language directions, Qwen reports better diarization accuracy than the competing systems it declines to name, plus gains in how faithful, fluent and concise the translations read. On the public FLEURS benchmark, covering 70 language directions, the company says it leads both its own prior generation and those same unnamed competitors on translation quality, latency, speech synthesis quality and speech recognition accuracy. The announcement names no rival system and publishes no scores, so outside developers have no way to check the size of the gap Qwen is claiming.
That gap matters more than it would in a slower-moving field, because live interpretation has become a feature race rather than a research corner. Google, Meta and OpenAI have each shipped some version of real-time spoken translation over the past two years, and each has made similar low-latency claims without a shared benchmark to test them against. What sets Qwen’s release apart is access: the model is already live through Alibaba’s DashScope API under the name qwen3.8-livetranslate-flash-realtime, which means a developer can run the 2.3-second lag and the diarization claims against their own audio rather than take the blog post’s word for it.
Qwen says its next targets are pushing latency down further, giving the model memory that persists across sessions for a recurring group of speakers, and extending coverage into more regional dialects. Teams evaluating live-translation vendors should run Qwen’s own audio through the DashScope endpoint before treating the 2.3-second figure as comparable to whatever number a competing vendor is currently quoting.
Reported from Qwen’s (Alibaba) official blog post announcing Qwen3.8-LiveTranslate, published September 18, 2026.
