Alibaba Qwen unveils Qwen3.8-LiveTranslate

The Qwen team wrote in its release post that simultaneous interpretation is not only about translating fast, and that it must also hear clearly and translate accurately. The same post said the team wants simultaneous interpretation to convey not only language but also the people and context behind a conversation. The official post on X was read more than 130,000 times within a day.

The architecture changed first. The approach, named Interleave, recasts simultaneous interpretation as a single stream in which audio and text are woven together in one line. Because audio already heard and translation already produced are cached and reused, quality rises while LAAL, the measure of average lagging, falls from 2.8 seconds in the previous generation to 2.3. The model is built from two modules, Thinker and Talker, on a Hybrid-MoE base. Thinker arranges video, audio, source text and translation into a single causal sequence in temporal order, and Talker combines the translation with the source audio to synthesize speech that preserves the speaker's timbre.

Three capabilities are new. The first is real-time speaker separation, which distinguishes speakers when several people talk in turn, attaches an owner to each sentence, and helps the translated speech hold each person's timbre more stably. The second is synchronized display of source and translation. Aligning the two languages side by side during interpretation supports immediate comprehension and source checking at once, and lays a foundation for follow-on features such as subtitle display, content organization and retrieval. The third is long-context disambiguation, which draws on earlier sentences and conversation history to clear up proper nouns and references and keep expression consistent.

Performance came in two evaluations. On Omnilingua-MSpeaker, which handles long audio with several speakers, the team said the model led current mainstream real-time interpretation systems across four dimensions covering 14 language directions: translation faithfulness, fluency and conciseness, plus diarization error rate. On the public FLEURS audio test set, it evaluated 70 language directions and said the model led both the previous generation and mainstream systems on translation quality, average lagging, speech recognition accuracy and speech synthesis quality.

Input and output coverage differ. Hearing audio and rendering it as text spans 60 languages, while speaking covers 29. The supported-language table on the Qwen blog, which METAL read in full, names them one by one, from Chinese, English, Korean and Japanese through Asturian, Cebuano and Tajik.

Using it comes down to a websocket. The real-time model is called qwen3.8-livetranslate-flash-realtime, and 16kHz PCM audio captured from a microphone goes to the server while 24kHz PCM speech comes back. Speaker separation and synchronized source output are switched on in the session configuration, and because the server returns the speaker identifier and the source text alongside the translation, there is no need to call a separate speech recognition interface. Where proper nouns are dense, hotwords can be registered to raise accuracy, and image frames can be sent so that what is on screen becomes a cue for the translation.

The demos are built from two dialogues from Journey to the West and a set of homophone examples. Three clips of 22, 28 and 12 seconds show speaker attribution and voice cloning, synchronized bilingual output, and how context and visual information resolve ambiguity.

Seen from the content side, the weight of this release sits on the name tag rather than the speed. Real-time subtitles have carried what was said while letting go of who said it, and in formats where speakers change often, such as panels and roundtables, a transcript alone did not let you rebuild the conversation. Once every sentence carries a speaker and the grain of each voice follows, live broadcasts, webinars and multilingual meeting records become usable source material without editing. Having source and translation on one screen also cuts review cost directly.

METAL has reported on OpenAI releasing a voice model that speaks while it listens, where the axis of competition was uninterrupted conversation. This release adds one more axis in the same real-time voice territory. Next to how fast the words come stands how precisely each one is attached to its speaker.

The Qwen team lists three directions ahead: compressing the gap between hearing and translating toward its limit, carrying memory across sessions for the same project and the same group of people, and extending coverage to long-tail languages and regional dialects. The unit that real-time interpretation handles is moving from the sentence to the conversation, and from the conversation to the relationship.