Filter Selection for Speaker Diarization Using Homomorphism

Abstract

In the current scenario, audio classification and recognition of a particular source of voice is a major challenge when several speakers are speaking at a time. The process of identifying each speaker in an audio segment is called speaker diarization. The major steps involved in speaker diarization are speech detection, speaker change, and speaker merges. Finding and suggesting the best filter is one of the most important task involved in every step of this process. No current researches at present have yet focused on the impact of filter with optimized approaches. In this chapter, a simple yet effective method using homorphism has been implemented to recommend the best filter for any audio classification task for this purpose. After having performed the classification task for a number of commonly used filters, their accuracies have been compared at every step of speaker diarization. In this work, a module using Bayesian information criterion, convolutional neural networks is implemented, and diarization algorithm that performs the task of speaker diarization is proposed.