Paper: https://arxiv.org/abs/2611.18330
Follow-up to last week's post, now with the paper.
You ask a typed question about an audio window (yes/no, a choice, or a score) and get a calibrated probability back. One frozen Qwen3-ASR-0.6B encoder pass per 8 s window serves all questions.
How each question type is answered: - Term ("was 'satellite' said?"): zero-shot, scored from the decoder's greedy and 4 beam hypotheses, then Platt-calibrated. No training on the terms. - Language: the ASR model's own language scores. - Turn end: a small readout on frozen layer 18. - Speaker change and count: readouts on layer 12.
Results: - On the same 1,000 public questions, it matches or beats Qwen3-Omni (30B-A3B) on 6 of 9 tests (image 1). Qwen3-Omni is still better on Speech Commands. Caveat: our turn-end readout was trained on Smart Turn data, so that row is in-distribution for us only. - Language ID: 100% on 300 FLEURS clips. - Speaker change AUROC 0.966-0.980 on AMI and VoxConverse, above ECAPA.…