Alibaba's Qwen team has introduced Qwen3.8-LiveTranslate, a new simultaneous interpretation system that rebuilds real-time translation around an Interleave architecture. According to the team, the design improves faithfulness, fluency, and conciseness while reducing average lagging (LAAL) from 2.8 seconds in the previous generation to 2.3 seconds.
The Interleave approach recasts simultaneous interpretation as a single audio-text interleaved stream. Both the audio already heard and the translation already produced can be cached and reused, which the team says markedly improves translation quality over the prior generation while lowering latency.
Building on support for 60 languages, Qwen3.8-LiveTranslate adds three capabilities aimed at real-world scenarios:
- Real-time speaker separation: the model distinguishes different speakers and what each says, with clear attribution for every sentence and more stable voice cloning that preserves each speaker's timbre.
- Synchronized source and translation: bilingual alignment on one screen supports instant comprehension and source checking, laying a foundation for subtitle display, content organization, and retrieval.
- Long-context disambiguation: by drawing on prior text and historical context, the model eases ambiguity around proper nouns and references, keeping expression consistent.
Under the hood, the system uses a Hybrid-MoE-based Thinker–Talker two-module design. The Thinker arranges video, audio, source text, and translation into a single causal sequence interleaved in temporal order and produced end to end, so understanding and translation happen within one sequence. The Talker then combines the translation and source audio to synthesize speech that preserves the original speaker's timbre.
On Omnilingua-MSpeaker, a multi-speaker long-audio evaluation set covering 14 language directions, Qwen3.8-LiveTranslate outperforms mainstream real-time interpretation systems across translation faithfulness, fluency, conciseness, and Diarization Error Rate (DER). On the public FLEURS audio test set across 70 language directions, it leads both the previous generation and current mainstream systems in translation quality, average lagging, speech recognition accuracy, and speech synthesis quality.
The team demonstrated speaker attribution and voice cloning, synchronized bilingual output, and context- and vision-assisted disambiguation using two dialogues from Journey to the West and a set of homophone examples. The new speaker separation and synchronized output features are enabled through session configuration via the DashScope API, with the server returning the speaker identifier and source-language text alongside the translation, without a separate ASR call.
Looking ahead, Qwen says it aims to compress end-to-end latency toward its limit, carry long-term memory across sessions for the same project and group of people, and extend coverage to more long-tail languages and regional dialects.