Qwen Releases LiveTranslate for Real-Time Interpretation
Alibaba’s Qwen team introduced a streaming interpretation model that combines speaker attribution, translation, and voice synthesis across dozens of languages.
Alibaba’s Qwen team has introduced Qwen3.8-LiveTranslate, a real-time interpretation model designed to process audio and visual context while producing translated text and speech.
The model uses a two-part Thinker–Talker architecture. The Thinker processes source audio, video, text, and translation in a single interleaved sequence, while the Talker generates translated speech and attempts to preserve the original speaker’s vocal identity. Qwen says the system supports 60 input languages and 29 spoken output languages, including Chinese, English, Japanese, Korean, Cantonese, Arabic, and Vietnamese.
Qwen reports results on two evaluation settings. On Omnilingua-MSpeaker, a long-audio, multi-speaker benchmark covering 14 language directions, it says LiveTranslate improves translation faithfulness, fluency, concision, and diarization error rate against mainstream real-time systems. On the FLEURS test set, covering 70 language directions, Qwen reports gains in translation quality, speech recognition accuracy, average lagging, and speech synthesis quality.
The model is available through Alibaba Cloud’s real-time API rather than as a public weight release. The commercial service accepts audio and images, returns text and audio, and is exposed through a WebSocket interface. Alibaba’s documentation lists a 2.3-second end-to-end latency for the Flash Realtime version.
The technical direction matters because simultaneous interpretation is constrained by several linked decisions: when to stop listening, how to preserve context across speakers, how to avoid translating names incorrectly, and how to produce speech before the full sentence is known. Qwen’s interleaved design tries to address those problems jointly instead of stitching together independent speech-recognition, translation, and text-to-speech systems.
Why it matters: the release strengthens China’s position in a commercially important multimodal category, though Qwen’s benchmark claims still need independent testing under noisy, overlapping, and low-resource language conditions.
Uncle Cat take
The significant detail is the 60-to-29 language path; Qwen is selling an interpretation pipeline, not merely a better speech translator.