Light Weight Mobile Device Targeted Speaker Clustering Algorithm

In this paper we present a novel light weight speaker clustering algorithm based on the Bayesian Information Criterion (BIC). Algorithm utilises BIC profiles, which were earlier used for False Alarm Compensation (FAC) in our Speaker Change Detector (SCD). Proposed speaker segmentation followed by a light weight clustering is targeted to segment and label mobile device recordings directly in the device itself. Thus the main criterion in algorithm design was to maintain high detection accuracy while keeping computational costs in low level. Clustering algorithm gave F-score performance of 0.90 for speaker segmentation, which is 29% relative improvement compared to baseline [1] results. Speaker segment labelling performance was 88%, when the number of speakers was undetermined. The experimental results indicate that our unsupervised speaker clustering algorithm is sufficiently effective and efficient for speaker segmentation applications in mobile devices.