Speaker Discrimination on Broadcast News and Telephonic Calls Using a Fusion of Neural and Statistical Classifiers

Abstract

This article describes a new Speaker Discrimination System (SDS), which is a part of an overall project called Audio Documents Indexing based on a Speaker Discrimination System (ADISDS). Speaker discrimination consists in checking whether two speech segments come from the same speaker or not. This research domain presents an important field in biometry, since the voice remains an important feature used at distance (via telephone). However, although some discriminative classifiers do exist nowadays, their performances are not enough sufficient for short speech segments. This issue led us to propose an efficient fusion between such classifiers in order to enhance the discriminative performance. This fusion is obtained, by using three different techniques: a serial fusion, parallel fusion and serial-parallel fusion. Also, two classifiers have been chosen for the evaluation: a mono-gaussian statistical classifier and a Multi Layer Perceptron (MLP). Several experiments of speaker discrimination are conducted on different databases: Hub4 Broadcast-News and telephonic calls. Results show that the fusion has efficiently improved the scores obtained by each approach alone. So, for instance, we got an Equal Error Rate (EER) of about 7% on a subset of Hub4 Broadcast-News database, with short segments of 4 seconds, and an EER of about 4% on telephonic speech, with medium segments of 10 seconds.