Qwen Releases 'Qwen3.8-LiveTranslate' with Real-Time Speaker Diarization and Source-Translation Synchronization

The Qwen team announced Qwen3.8-LiveTranslate, which enhances the accuracy and listening capabilities of real-time simultaneous interpretation. The model introduces an Interleave architecture that restructures audio and text into a single interleaved stream, reducing average latency (LAAL) from 2.8 seconds to 2.3 seconds while improving translation fidelity and fluency.

Key Features and Architecture

The model adopts a Thinker-Talker 2-module design based on Hybrid-MoE. The Thinker processes video, audio, source text, and translation as a single causal sequence, while the Talker synthesizes speech by combining translation with original audio to stably preserve the speaker's timbre. Key new features include real-time speaker diarization, synchronized output of source and translated text, and Long-context disambiguation.

Performance and Language Support

The model outperformed existing systems in translation quality and speaker diarization error rate (DER) on the Omnilingua-MSpeaker benchmark, and led in speech recognition and synthesis quality across 70 language directions on the FLEURS benchmark. Input audio and output text support 60 languages, while output audio supports 29 languages, including Korean, English, and Chinese.

API and Usage

By streaming microphone audio via the DashScope API, the server returns speaker identifiers and source text along with the translation, eliminating the need for separate ASR interface calls. Session settings allow enabling speaker diarization and synchronized output, and the system also supports hotword registration for improved proper noun accuracy and visual context transmission.