New Diagnostic Ladder Separates Decision-Rule and Readout-Coverage Gaps in Speech Language Models
A new study on arXiv (2608.06409) introduces a diagnostic ladder that helps evaluate speech language models in tasks related to paralinguistics, focusing on generation. The method looks at the generated responses, option logits, and different readouts of these logits and hidden states to pinpoint where there are shortcomings in endpoints and decision-making processes. When examining five systems across two emotion datasets, it turns out that state decoding is about 27.8 points more accurate than generation. Additionally, a logit correction technique improves accuracy across the board, hinting that some decision-making issues can be fixed. The research suggests that emotional insights can apply to new speakers and situations. This paper can be found online.
Key facts
- Paper arXiv:2608.06409 introduces a generation-aligned diagnostic ladder for speech language models.
- The ladder compares emitted answer, option logits, affine readout, and linear readout of hidden state.
- State decoding exceeds generation by 27.8 accuracy points on average across five systems and two emotion corpora.
- Decision-rule and readout-coverage gaps are positive in all ten conditions tested.
- A label-free logit correction improves generated accuracy in every condition.
- Emotion information outside the native readout generalizes to held-out speakers and scenarios.
- The paper is a cross-type announcement on arXiv.
- The study focuses on paralinguistic tasks in speech language models.
Entities
Institutions
- arXiv