How to Deal with Multiple-Targets in Speaker Identification Systems?

Abstract

In open-set speaker identification systems a known phenomenon is that the false alarm (accept) error rate increases dramatically when increasing the number of registered speakers (models). In this paper, we demonstrate this phenomenon and suggest a solution using a new modeldependent score-normalization technique, called Top-norm. The Top-norm method was specifically developed to improve results of open-set speaker identification systems. Also, we suggest a score-normalization parameter adaptation technique. Experiments performed using speaker recognition corpora are described and demonstrate that the new method outperforms other normalization methods.

The effect of room parameters on speaker verification using reverberant speech

“…When T 60 = 0. 5 of n seem to be small. However, CMS performance may not be explained entirely by MDR values.…”

“…The presence of reverberation adds distortion to the feature vectors, which results in performance degradation of SVR and thus of speaker recognition, due to mismatched conditions between trained models and test segments [3]. Feature normalization techniques such as cepstral mean subtraction (CMS) and score normalization techniques such as the Znorm, Hnorm, Tnorm [4] and Top-norm [5] may be used to reduce the effect of reverberation which is characterized by a short-duration room impulse response (RIR). However, they might not be as useful to reduce the effect of long-duration RIR, which is often the case in room acoustics [6].…”

MCE 2018: The 1st Multi-Target Speaker Detection and Identification Challenge Evaluation

“…Multitarget detection such as blacklist or watchlist was often described as open-set speaker identification. There are a few relevant studies [2,3,4,5,6,7], but it is not actively being explored because it is regarded as a special case of speaker verification. Most research on this topic pre-date the i-vector [8], so it is difficult to compare the performance of older blacklist detection systems with state-of-the-art technology.…”

The Effect of Reverberation on Optimal GMM Order and CMS Performance in Speaker Verification Systems

“…The presence of reverberation adds distortion to the feature vectors, which results in performance degradation of SVR systems due to mismatched conditions between trained models and test segments. Feature normalization techniques such as the cepstral mean subtraction (CMS) [Mammone et al, 1996] and variance normalization [Chen & Bilmes, 2007], and score normalization techniques such as the Znorm, Hnorm, Tnorm [Bimbot et al, 2004, Mammone et al, 1996 and Top-norm [Zigel & Wasserblat, 2006], were originally developed to compensate for the effect of a telephone channel [Mammone et al, 1996], or for the effect of slowly varying convolutive noises in general [Reynolds et al, 2000]. For that reason, these techniques may be used to reduce the effect of reverberation, if it is characterized by a short-duration room impulse response (RIR).…”