Text-independent speaker verification by discriminator counting

Abstract

This paper describes an algorithm for detecting the presence of speech from a particular individual, called the target, while rejecting speech from all other individuals. The algorithm requires samples of training data for the target speaker, as well as for a set of N other speakers, called reference speakers. This data is used to estimate the parameters of a set of speaker-pair discriminators. Each discriminator separates the target from one of the reference speakers, producing a positive or negative value for each input frame. The polarity of discriminators is such that positive output values favor the target. After integrating the discriminator outputs over a suitable time interval, the number of positive values is counted and compared with a detection threshold between 0 and N. In an experiment using clean speech materiai, 80% of targets were detected with only 2% of non-targets. INTRODUCTIONOther speaker verification algorithms have attempted to evaluate the likelihood that the observed data would be produced by the target speaker, without consideration of the likelihood that the data would be produced by other speakers. This approach has led to high variability of performance from one target speaker to another, and lack of robustness with respect to message content and channel conditions. Recently, an algorithm based on estimation of the likelihood ratio[Higg, Higg891 was developed to address this problem. The likelihood-ratio approach compares match scores derived from the target model with match scores derived from the models of other speakers (the reference speaker set). The method described here differs from previous methods, including the likelihood-ratio approach, by modeling speaker-pair distinctions, rather than speakers themselves. The potential advan-

Topic and speaker identification via large vocabulary continuous speech recognition

“…Several approaches to the problems of topic and speaker identification have already appeared in the literature. For example, an approach to topic identification using wordspotting is described in [1] and approaches to the speaker identification problem are reported in [2] and [3]. Dragon Systems' approach to the message identification tasks depends crucially on the existence of a large vocabulary continuous speech recognition system.…”

A novel text-independent speaker verification method based on the global speaker model

“…Research on speaker verification has been focused on speaker models [2], feature selection [3], and robust methods [4]. Higgins used a discriminate counting to verify the speaker [5], as well as likelihood score normalization methods [6]- [8], which are two likelihood score normalization methods by using impostor models. Since the method in [7] improves the speaker verification rate over the method in [6], it is used in this correspondence as a comparison method.…”