The Impact of Audio Segmentation to Speaker Tracking in Broadcast News Data

A system for speaker tracking in broadcast-news audio data is presented. The process of speaker tracking in continuous audio streams involves several processing tasks and is therefore treated as a multistage process. The main building blocks of such system include the components for audio segmentation, speech detection, speaker clustering and speaker identification. Our system was developed by implementing the most recent published methods in each component of the system, whereas we focused mainly on the component for audio segmentation for being considered as one of the most critical components of such systems. Two alternative approaches of speaker change detection and speaker clustering are explored and their impacts to the overall speaker-tracking performance are evaluated. The evaluation experiments were performed on broadcast-news audio data with a speaker-tracking system capable of detecting 41 target speakers. The comparison of the evaluation results of different versions of the speaker-tracking system indicates the importance of the tasks in the audio-segmentation module and provides valuable insights into how the system works.