Marimba: A System One Model for Speech
Marimba uses a frozen speech recognizer to answer questions about words, speakers, language and whether a turn has ended.
Built on Qwen3-ASR 0.6B features, it combines transcript scoring with specialized readouts for non-lexical decisions. Speaker-change detection reached AUROC 0.966–0.980 on AMI and VoxConverse, outperforming an ECAPA baseline; speaker counting also beat pyannote 3.1. Turn-end detection scored 0.930 against Smart Turn v3.2’s 0.982 on that system’s own test set. On an ICSI meeting subset, the models scored 0.716 and 0.712, with no statistically clear difference.
HF Daily Papers' note
Built on Qwen3-ASR 0.6B features, it combines transcript scoring with specialized readouts for non-lexical decisions. Speaker-change detection reached AUROC 0.966–0.980 on AMI and VoxConverse, outperforming an ECAPA baseline; speaker counting also beat pyannote 3.1. Turn-end detection scored 0.930 against Smart Turn v3.2’s 0.982 on that system’s own test set. On an ICSI meeting subset, the models scored 0.716 and 0.712, with no statistically clear difference.
HF Daily Papers' note
score 4