Support vector machines for speaker based speech indexing

This paper proposes an integrated framework for speaker indexing which includes both speaker segmentation and speaker clustering. Speaker indexing systems has wide domains of application with different requirements which make a general speaker indexing framework hard to accomplish. The main source of performance degradation in speaker indexing is the probable existence of short speech utterances which makes the speaker turns hard to distinguish and also exposes the segment modeling to data insufficiency. This paper introduces a speaker indexing framework with high average performance which uses Support Vector Machines (SVM) as the core approach. The main contribution of this framework is the SVM based clustering approach which makes the indexing more robust against the short speech segments. This framework is evaluated on a domestic conversational speech dataset and the results were satisfactory.