Alibaba's Qwen Team Unveils Qwen3.8-LiveTranslate: A Real-Time Speech Translation Model Achieving Just 2.3 Seconds Average Latency in 60 Languages

Qwen has unveiled Qwen3.8-LiveTranslate, a new model built for real-time simultaneous interpretation. The system listens to live speech—optionally paired with video frames—and delivers translated text and audio while the original speaker is still talking. The headline improvement is a redesigned Interleave architecture, which Qwen says boosts faithfulness, fluency, and conciseness while cutting average lagging (LAAL) from 2.8 seconds down to 2.3 seconds. The launch also brings live speaker diarization, side-by-side bilingual display, and long-context disambiguation. Developers can access it today as a hosted API through Alibaba Cloud Model Studio and QwenCloud under the ID qwen3.8-livetranslate-flash-realtime over WebSocket.

What Changed Under the Hood

Simultaneous interpretation always involves a compromise: waiting longer gives the model richer context, but responding faster reduces delay for listeners. Qwen3.8-LiveTranslate reworks this loop with its new Interleave architecture.

The latency metric here is LAAL, or Length-Adaptive Average Lagging, which measures how far the translation trails behind the source speech while avoiding rewards for systems that simply over-generate output. Moving from 2.8 seconds to 2.3 seconds represents roughly an 18% reduction in average lag.

According to the QwenCloud team, this model is the real-time version of Qwen3.8-LiveTranslate-Flash. It builds on the Qwen-Omni stack, large-scale multimodal training data, cross-language and cross-modal alignment, plus visual enhancement. The Flash variant also handles offline audio and video translation.

Three New Capabilities

Real-time speaker diarization: The model separates individual speakers in multi-party conversations and preserves each person's voice through more stable voice cloning. The API exposes cloning modes, including an "always" mode that re-clones before each response for multi-speaker sessions.

Synchronized bilingual display: Source text and its translation appear on screen together. In the API, the source transcription streams as separate events alongside the translation stream.

Long-context disambiguation: The model draws on conversation history to resolve names and terminology, so a name introduced early in a meeting stays consistent throughout the translation.

Languages, Inputs, and Vision

The model comprehends 60 languages and can speak 29 of them, returning audio with text; the other 31 are text-only. Speech output spans Chinese, English, Arabic, German, French, Spanish, Japanese, Korean, Hindi, and more.

Inputs consist of audio and optional images, with outputs in text and audio. Visual cues—lip movements, gestures, on-screen text—help in noisy environments and with ambiguous words. Documentation advises sending no more than 2 images per second. Teams can also configure hotwords mapping source terms to fixed translations, with a recommended cap of 1,000 entries.

API, Pricing, and Limits

Default audio runs at 16 kHz PCM input and 24 kHz PCM output, with Tina as the default voice. Clients can set session.output_modalities to text-only or text plus audio, and should always send session.finish before closing or the final segment is lost.

Singapore list pricing per 1M tokens: audio input $7.50, image input $0.55, text output $20, audio output $30. Beijing pricing is lower at $5.653, $0.466, $14.133, and $22.613 respectively. Audio input consumes 7 tokens per second and audio output 12.5 tokens per second; one hour of two-way speech costs roughly $1.54 in Singapore, excluding text and image tokens.

The context window spans 53,248 tokens (49,152 input, 4,096 output), with default rate limits of 10 requests and 100,000 tokens per minute. Model Studio lists function calling, structured outputs, batch inference, and fine-tuning as unsupported.

What to Watch

Qwen3.8-LiveTranslate lowers average lag by nearly a fifth while adding diarization, bilingual display, and context-aware disambiguation across 60 languages—though access remains API-only through Alibaba's platforms. It remains to be seen how competitors in live translation respond, and whether open weights or offline deployment options follow.

Meta description: Qwen3.8-LiveTranslate delivers real-time speech translation across 60 languages, cutting average lagging from 2.8s to 2.3s with a new Interleave architecture.

Tags: Qwen, Qwen3.8-LiveTranslate, real-time translation, Alibaba Cloud, speech-to-speech

Featured image: Abstract technology style illustration of overlapping audio waveforms flowing between two color-coded streams, no people, logos, or text.