A study of speaker clustering for speaker attribution in large telephone conversation datasets

摘要

This paper proposes the task of speaker attribution as speaker diarization followed by speaker linking. The aim of attribution is to identify and label common speakers across multiple recordings. To do this, it is necessary to first carry out diarization to obtain speaker-homogeneous segments from each recording. Speaker linking can then be conducted to link common speaker identities across multiple inter-session recordings. This process can be extremely inefficient using the traditional agglomerative cluster merging and retraining commonly employed in diarization. We thus propose an attribution system using complete-linkage clustering (CLC) without model retraining. We show that on top of the efficiency gained through elimination of the retraining phase, greater accuracy is achieved by utilizing the farthest-neighbor criterion inherent to CLC for both diarization and linking. We first evaluate the use of CLC against an agglomerative clustering (AC) without retraining approach, traditional agglomerative clustering with retraining (ACR) and single-linkage clustering (SLC) for speaker linking. We show that CLC provides a relative improvement of 20%, 29% and 39% in attribution error rate (AER) over the three said approaches, respectively. We then propose a diarization system using CLC and show that it outperforms AC, ACR and SLC with relative improvements of 32%, 50% and 70% in diarization error rate (DER), respectively. In our work, we employ the cross-likelihood ratio (CLR) as the model comparison metric for clustering and investigate its robustness as a stopping criterion for attribution.