On the calibration of powerset speaker diarization models
This work addresses calibration issues in speaker diarization for audio processing, offering incremental improvements in model reliability and data selection.
The paper tackles the calibration of powerset speaker diarization models, showing that top-label confidence reliably predicts high-error regions and that training on low-confidence data improves model calibration and annotation efficiency.
End-to-end neural diarization models have usually relied on a multilabel-classification formulation of the speaker diarization problem. Recently, we proposed a powerset multiclass formulation that has beaten the state-of-the-art on multiple datasets. In this paper, we propose to study the calibration of a powerset speaker diarization model, and explore some of its uses. We study the calibration in-domain, as well as out-of-domain, and explore the data in low-confidence regions. The reliability of model confidence is then tested in practice: we use the confidence of the pretrained model to selectively create training and validation subsets out of unannotated data, and compare this to random selection. We find that top-label confidence can be used to reliably predict high-error regions. Moreover, training on low-confidence regions provides a better calibrated model, and validating on low-confidence regions can be more annotation-efficient than random regions.