← News

Models & toolsSources: arXiv · 16 Nov 2026

Marimba paper describes typed speech decisions built on frozen Qwen3-ASR features

An arXiv abstract reports a system that answers typed questions about speech using frozen ASR features, with measured accuracy on private and public test sets.

In plain words

Researchers describe Marimba, a system that makes specific decisions about audio, such as whether a word was said, whether a speaker has finished their turn, or which language is being spoken. It reuses features from an existing speech recognition model rather than training a new one from scratch. The reported accuracy comes from the paper's own abstract and has not been independently verified.

What was announced

The abstract presents Marimba as an interface for typed speech decisions built on frozen Qwen3-ASR 0.6B features. Lexical questions are answered by zero-shot scoring of ASR transcripts, non-lexical questions use a question-conditioned head, and language identification relies on the ASR model's own language scores. On private conversational test sets, the authors report decoder AUROC of 0.80-0.92, rising to 0.92-0.97 on public read speech, with Speech Commands four-way accuracy of 0.93 and binary lexical ECE of at most 0.08.

The authors also report a personal-information detection pipeline that reaches item-level F1 of 0.904 on synthetic Brazilian Portuguese speech, and layer-12 speaker readouts that detect speaker change with AUROC of 0.966-0.980 on AMI and VoxConverse, above an ECAPA baseline. The turn-end readout scores AUROC 0.930 against Smart Turn v3.2 at 0.982 on Smart Turn's own test set. On a subset of ICSI meetings, the difference from a comparison point is within the paired confidence interval, which the abstract notes includes zero.

What it does not show

  • Evidence is limited to the paper's abstract; the full article was not reviewed, so methods and caveats beyond the abstract are unverified.
  • All reported figures are the authors' own measurements on their chosen test sets, including private and synthetic data, and no independent confirmation is available.
  • The abstract states the turn-end comparison confidence interval on ICSI meetings includes zero, so that result is not established as an improvement.

For security and IT teams

Scope
The reported results cover term presence, turn completion, language identification, PII detection, speaker change and speaker counting, evaluated on private conversational sets, public read speech, synthetic Brazilian Portuguese, AMI, VoxConverse, Smart Turn and ICSI data.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.