Qwen3.8-LiveTranslate adds speaker separation

Alibaba's Qwen team has released Qwen3.8-LiveTranslate, a new version of its real-time interpretation model. As we reported earlier, the system already supported 60 languages. This release keeps that coverage and adds three features: speaker separation, source text and translation shown together on one screen, and the use of earlier conversation to translate names and terms more accurately.

The main change is what Qwen calls an Interleave architecture. Video, audio, source text and translation are placed in a single sequence in time order, so the model can cache and reuse both the audio it has heard and the translation it has already produced. Qwen says this improves faithfulness, fluency and conciseness, and lowers average lagging (LAAL) from 2.8 seconds to 2.3 seconds.

The model uses a Hybrid-MoE "Thinker–Talker" design. The Thinker handles understanding and translation. The Talker turns the translation into speech that keeps each original speaker's voice, and speaker separation is meant to make that voice cloning more stable when several people take turns.

On Omnilingua-MSpeaker, a multi-speaker long-audio test set covering 14 language directions, Qwen says the model beats "mainstream" real-time interpretation systems on faithfulness, fluency, conciseness and diarization error rate. On FLEURS, across 70 language directions, it claims to lead on translation quality, lag, speech recognition accuracy and speech synthesis quality. The post does not name the competing systems or give their scores.

The model accepts audio in 60 languages and can speak translations in 29. It is available through Alibaba's DashScope API. Speaker IDs and source-language text come back alongside the translation, so developers do not need a separate speech-recognition call. The post does not mention pricing or open weights.