Model releases & benchmarks
Qwen3.8-LiveTranslate: Names the speaker. Carries the meaning.
Qwen3.8-LiveTranslate combines interleaved audio and text with speaker separation, bilingual alignment, and context across turns; the article reports LAAL falling from 2.8 seconds to 2.3 seconds.
Why it matters: Speaker labels and aligned source text can help users follow who said what and check a translation as it arrives. The language counts also matter in practice: the listed output-audio coverage is narrower than input-audio/output-text coverage.