Qwen Releases Qwen3.8-LiveTranslate, a Real-Time Simultaneous Interpretation Model with 2.3-Second Latency

The Qwen team has released Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model that processes audio and text as a single Interleave stream. This model adopts a Hybrid-MoE-based Thinker-Talker dual-module structure and synthesizes translated speech while preserving the original speaker's timbre.

The key performance metric, average latency (LAAL), has decreased from 2.8 seconds in the previous generation to 2.3 seconds. On the Omnilingua-MSpeaker benchmark (14 language directions), it demonstrated advantages over major competing systems in four categories: translation fidelity, fluency, conciseness, and speaker diarization error rate (DER).

Supported languages include 60 for input audio and text, and 29 for output audio, with Korean included in all categories. It provides real-time speaker diarization, synchronized source-translation output, and ambiguity resolution based on long-term context. It is available via the DashScope API under the model name qwen3.8-livetranslate-flash-realtime.