Alibaba Qwen releases Qwen3.8-LiveTranslate with 2.3s latency across 60 languages
On September 19, Alibaba's Qwen team released Qwen3.8-LiveTranslate, a real-time simultaneous interpretation model covering 60 languages, cutting average latency (LAAL) from 2.8s to 2.3s. It promises a listen-translate-speak experience and is available via the streaming API qwen3.8-livetranslate-flash-realtime on Alibaba Cloud Bailian.
Confirmed
- Built on an Interleave architecture that reframes simultaneous interpretation as a single interleaved audio-text stream, with caching and reuse of both heard audio and emitted translations; per @xiaohu, the design includes a Thinker-Talker structure
- Improves faithfulness, fluency and conciseness across 60 languages
- Supports streaming speaker diarization (spk1/spk2/spk3 tags), with translated speech preserving each speaker's voice timbre
- Supports synchronized bilingual subtitles and long-context understanding of names and terminology
Why it matters
- The 2.3s latency brings real-time simultaneous interpretation closer to practical use in meetings and multi-party conversations, while speaker diarization and voice preservation cover real dialogue needs
- The streaming API on Alibaba Cloud Bailian lets developers directly build real-time translation applications
2026-09-19 ~ 2026-09-21 · 6 related posts
Primary sources
- Qwen ships LiveTranslate real-time interpreter with lag cut to 2.3s across 60 languages — Alibaba_Qwen ·
- Alibaba releases Qwen3.8-LiveTranslate: simultaneous interpretation model with 2.3s latency — 千问大模型 ·
- Alibaba's Qwen3.8-LiveTranslate does real-time speech translation with 2.3s lag and speaker separation — xiaohu ·
- [source] Alibaba releases Qwen3.8-LiveTranslate: simultaneous interpretation model with 2.3s latency — 千问大模型 · 2026-09-19
- [source] Qwen ships LiveTranslate real-time interpreter with lag cut to 2.3s across 60 languages — Alibaba_Qwen · 2026-09-19
4 near-duplicate retellings: airesearch12 · xiaohu · xiaohu · emmanuelvivier