Speaker diarization of news broacasts and meeting recordings

Abstract

Given a piece of audio recording, the task of speaker diarization can be summarized as answering the question of "Who spoke when ?". This thesis offers a review of the techniques and issues relating to performing speaker diarization on broadcast news recordings, as well as meeting recordings. The broadcast news domain is generally regarded to be simpler because the turn taking between speakers is better controlled and audio quality tends to be higher. The typical approach used for this domain consist of two steps-speaker segmentation and then speaker clustering. The Bayesian Information Criterion (BIC) has been a very popular distance measure for both speaker segmentation and clustering. Experiments were conducted that confirmed the effectiveness of this distance measure for segmentation and clustering. Further speaker segmentation experiments were performed using the Hotelling's T 2 statistic to augment the BIC. It was observed that while this does speed up processing, the segmentation F Score obtained does not match up to that reported in the literature. A novel speaker clustering approach was also introduced where polynomial expanded feature vectors were used to compute the distance between clusters. It was found that this approach could produce results comparable to that for the BIC. In order to address the problem of speaker diarization for the meeting domain, a diarization system was developed and submitted for the NIST Rich Transcription 2007 (RT-07) evaluation. This diarization system exploited the diversity of meeting recording channels by performing Time Delay of Arrival (TDOA) estimation using a Normalized Least Means Squared (NLMS) filter. Subsequent performance enhancements were delivered by adding a cluster purification module, as well as a Non-Speech & Silence Removal (NS&SR) module. An overall Diarization Error Rate (DER) of 15.32% was obtained for the RT-07 corpus. This score was found to be competitive against the other entrants in the evaluation exercise.

Fusing eye-gaze and speech recognition for tracking in an automatic reading tutor – a step in the right direction?

“…The evaluation method used in this paper builds on [6] but moves away from the notion of speech events (segments with either a word or miscue) to word index changes -that is whenever the focus moves from one word to another. Since the tracking reference was generated by a human who applies judgment when placing changing points, a tolerance interval is introduced similar to that in [12]. Another reason for using a tolerance interval is that even though the reading tutor might be slightly off in detecting a changing point this has no practical significance in most settings.…”