Qwen3.8-LiveTranslate: speaker-aware simultaneous interpretation

Qwen3.8-LiveTranslate adds real-time speaker separation, aligned source and translation text, and contextual disambiguation. Its 18 September announcement describes an interleaved audio-text architecture and a DashScope real-time API example. Qwen reports average lagging of 2.3 seconds in its evaluation.

Language support is uneven across outputs. The announcement lists 60 languages for input audio and output text, including Afrikaans, and 29 for output audio. Afrikaans is absent from the spoken-output list. Check the exact direction and output format your workflow requires before describing the product as supporting spoken translation for every listed language.

Use consented recordings with known translations and clear speaker labels. Include names, specialised terminology and people taking turns quickly. Have a fluent reviewer assess whether meaning and attribution survive. Record mistakes separately: an accurate sentence assigned to the wrong speaker can be a serious problem in meeting notes.

Then evaluate a live conversation under your actual network conditions. Measure the delay a listener experiences and the system's response to interruptions. Keep the original-language transcript available alongside the translation so a reviewer can inspect uncertain passages. Confirm account region and service availability before designing a dependency around the example endpoint. For consequential conversations, decide when a human interpreter must take over before the pilot begins.