Overview
Qwen3.8-LiveTranslate is a real-time speech interpretation model released by Alibaba's Qwen team on September 18, 2026. It reframes simultaneous interpretation as a single interleaved stream of audio and text: audio already heard and translations already produced are cached and reused, with understanding and translation happening inside one sequence. The company puts the average interpretation lag (LAAL) at 2.3 seconds, down from 2.8 seconds in the previous generation.
On language coverage, the model takes audio input in 60 languages and produces translated text, with spoken output in 29 of them that preserves the original speaker's voice. This generation adds three capabilities aimed at real meeting scenarios: real-time speaker diarization that tags each translated sentence with its speaker, on-screen alignment of source and translated text for bilingual reading, and long-context disambiguation that uses prior history to keep names and terminology consistent across turns.
On benchmarks, the company reports that on Omnilingua-MSpeaker, a long-form multi-speaker evaluation across 14 language pairs, the model leads current mainstream real-time interpretation systems on fidelity, fluency, conciseness and speaker diarization error rate, and that on the public FLEURS test set, covering 70 language pairs, it leads the previous generation and mainstream systems on translation quality, lag, recognition accuracy and synthesis quality. The model is served through Alibaba Cloud Bailian's realtime API under the name qwen3.8-livetranslate-flash-realtime; no open weights were published at the time of collection.
Key Features
- Single-stream Interleave architecture: Audio and translation are generated interleaved in one causal sequence, with heard audio and produced translation cached and reused, improving quality and lag together.
- Thinker-Talker design: The Thinker arranges video, audio, source text and translation into a single sequence for streaming understanding and translation, while the Talker synthesizes speech that preserves the original speaker's voice.
- Real-time speaker diarization: The server returns speaker identifiers and source-language text alongside each translation, with no separate ASR pipeline needed, keeping attribution clear in multi-speaker meetings.
- Bilingual on-screen alignment: Source and translated text stream out in sync, ready for subtitles, meeting minutes and interpreter support, enabled through session configuration.
- Long-context disambiguation: Links history across turns to keep names and terminology consistent, with hotword registration and image-frame input available as additional context.
- 60-language input, 29-language voice output: Text translation covers 60 languages including Chinese, English, Japanese, Korean, German, French, Spanish, Russian, Arabic and Cantonese; 29 of them can be spoken back with preserved voice.
Use Cases
- Live subtitles for international conferences and streams: low-latency bilingual captions with speaker attribution
- Cross-language support and voice agents: understand and translate in real time, speak back in the original voice
- Meeting minutes and interview processing: get source text, translation and speaker labels in one pass
- Interpreter assist tools: low-latency draft translations for human interpreters to verify
- Multilingual video production: first-pass translation and alignment for dubbing and subtitles
Pros
- The company's LAAL figure of 2.3 seconds sits within usable range for live interpretation
- Speaker diarization and bilingual alignment are built into session configuration, no custom pipeline required
- Spoken output preserves the original speaker's voice, which keeps multi-speaker scenarios natural
- Same generation as the Qwen3.8 family, so mixing with the Omni line costs little adaptation
- Text translation spans 60 languages, a wide spread for real-time interpretation
Pricing
Served through Alibaba Cloud Bailian's realtime API as qwen3.8-livetranslate-flash-realtime, billed by audio duration and output volume. The launch post does not publish pricing details; confirm on the Bailian platform.
Summary
Qwen3.8-LiveTranslate is a real-time interpretation model released by Alibaba's Qwen team on September 18, 2026, using a single interleaved stream architecture to reach a company-reported LAAL of 2.3 seconds, with 60-language text input and 29-language voice output. Speaker diarization, bilingual alignment and long-context disambiguation are all built into the session configuration. Teams running live conference captions, cross-language voice agents or interpreter assist can plug in directly through Bailian's realtime API.
Version History
- Qwen releases the Qwen3.8-LiveTranslate real-time simultaneous interpretation model, with LAAL reduced to 2.3 seconds. (2026-09-18): Qwen releases Qwen3.8-LiveTranslate, adopting an Interleave architecture and a Hybrid-MoE Thinker-Talker design to rebuild real-time simultaneous interpretation, reducing the average lag (LAAL) from 2.8 seconds in the previous generation to 2.3 seconds.