Alibaba's Qwen team launches simultaneous translation model cutting average latency to 2.3 seconds

Alibaba's Qwen team has released Qwen3.8-LiveTranslate, a real-time interpretation model that listens to live speech, with or without video frames, and returns translated text and audio while the speaker is still talking. The core change is an Interleave architecture that rebuilds the loop between waiting for context and early exit. The Length-Adaptive Average Lagging metric, LAAL, fell from 2.8 seconds to 2.3 seconds, an improvement of roughly 18 percent in average latency. The model is available as a hosted API on Alibaba Cloud Model Studio and QwenCloud under the identifier qwen3.8-livetranslate-flash-realtime over WebSocket.

The model addresses three practical pain points in live translation. Real-time speaker diarization distinguishes participants in multi-voice conversations and preserves each speaker's voice through more stable cloning; the API exposes cloning modes, including an "always" mode that re-clones before every response in multi-speaker sessions. A synchronized bilingual display presents source text and translation side by side, with transcription flowing as separate events alongside the translation stream. Long-context ambiguity resolution uses conversation history to maintain name and term consistency, so a name mentioned at the start of a meeting remains identical in translation even an hour later.

The model understands 60 languages and can produce speech in 29, including Chinese, English, Arabic, German, French, Spanish, Japanese, Korean and Hindi; the remaining languages return text only. Accepted input includes audio and optional images, up to two images per second recommended. Visual cues such as lip movements, gestures and on-screen text help in noisy environments and with ambiguous words. Teams can define hotwords, mapping source terms to fixed translations, with a limit of 1,000 mappings per session.

Connection uses a WebSocket Realtime API with speaker identification enabled by default; the client streams continuous audio and receives server responses. Input audio runs at 16 kHz PCM, output at 24 kHz PCM. The default voice is Tina, and output can be limited to text only. Clients must send session.finish before closing the connection, otherwise the final segment is lost. The context window stands at 53,248 tokens, split between 49,152 for input and 4,096 for output.

Pricing in Singapore: $7.50 per million audio input tokens, $0.55 per image, $20 per million text output tokens, $30 per million audio output tokens. Beijing pricing is roughly 25 percent lower. Token consumption runs at 7 per second of audio input and 12.5 per second of audio output. One hour of two-way speech costs approximately $1.54 in Singapore before text and image tokens.