Novel Approaches to Speaker Clustering for Speaker Diarization in Audio Broadcast News Data

Abstract

The growing demand to shift content-based information retrieval from text to various multimedia sources means there is an increasing need to deal with large amounts of multimedia information. The data provided from television and radio broadcast news (BN) programs are just one example of such a source of information. In our research we focus on the processing and analysis of audio BN data, where the main information source is represented by speech data. The main issues in our work relate to the preparation and organization of BN audio data for further processing in information audio-retrieval systems based on speech technologies. This chapter addresses the problem of structuring the audio data in terms of speakers, i.e., finding the regions in the audio streams that belong to a single speaker and then joining each region of the same speaker together. The task of organizing the audio data in this way is known as speaker diarization and was first introduced in the NIST project of Rich Transcription in the “Who spoke when” evaluations (Fiscus et al., 2004; Tranter & Reynolds, 2006). The speaker-diarization problem is composed of several stages, in which the three main tasks are performed: speech detection, speakerand background-change detection, and speaker clustering. While the aim of the speech detection and the speakerand acousticsegmentation procedures is to provide the proper segmentation of the audio data streams, the purpose of the speaker clustering is to join or connect together segments that belong to the same speakers, and this is usually applied in the last stage of the speaker-diarization process. In this chapter we focus on speaker-clustering methods, concentrating on developing proper representations of the speaker segments for clustering, and research different similarity measures for joining the speaker segments and explore different stopping criteria for the clustering that result in a minimization of the overall diarization error of such systems. The chapter is organized as follows: In Section 2, two baseline speaker-clustering approaches are presented. The first is a standard approach using a bottom-up agglomerative clustering principle with the Bayesian information criterion as the merging criterion. In the second system an alternative approach is applied, also using bottom-up clustering, but the representations of the speaker segments are modeled by Gaussian mixture models, and for O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg