Search

In this paper we propose the use of Support Vector Machine Regression (SVR) for robust speaker verification in two scenarios: i) strong mismatch in speech conditions and ii) forensic environment. The proposed approach seeks robustness to situations where a proper background database is reduced or not present, a situation typical in forensic cases which has been called database mismatch. For the mismatching condition scenario, we use the NIST SRE 2008 core task as a highly variable environment, but with a mostly representative background set coming from past NIST evaluations. For the forensic scenario, we use the Ahumada III database, a public corpus in Spanish coming from real authored forensic cases collected by Spanish Guardia Civil. We show experiments illustrating the robustness of a SVR scheme using a GLDS kernel under strong session variability, even when no session variability is applied, and especially in the forensic scenario, under database mismatch. [ABSTRACT FROM AUTHOR]

In this work, we analyze several quality measures for speaker verification from the point of view of their utility, i.e., their ability to predict performance in an authentication task. We select several quality measures derived from classic indicators of speech degradation, namely ITU P.563 estimator of subjective quality, signal to noise ratio and kurtosis of linear predictive coefficients. Moreover, we propose a novel quality measure derived from what we have called Universal Background Model Likelihood (UBML), which indicates the degradation of a speech utterance in terms of its divergence with respect to a given universal model. Utility of quality measures is evaluated following the protocols and databases of NIST Speaker Recognition Evaluation (SRE) 2006 and 2008 (telephone-only subset), and ultimately by means of error-vs.-rejection plots as recommended by NIST. Results presented in this study show significant utility for all the quality measures analyzed, and also a moderate decorrelation among them. [ABSTRACT FROM AUTHOR]

In this work, a novel method for system fusion in emotion recognition for speech is presented. The proposed approach, namely Anchor Model Fusion (AMF), exploits the characteristic behaviour of the scores of a speech utterance among different emotion models, by a mapping to a back-end anchor-model feature space followed by a SVM classifier. Experiments are presented in three different databases: Ahumada III, with speech obtained from real forensic cases; and SUSAS Actual and SUSAS Simulated. Results comparing AMF with a simple sum-fusion scheme after normalization show a significant performance improvement of the proposed technique for two of the three experimental set-ups, without degrading performance in the third one. [ABSTRACT FROM AUTHOR]

A new paradigm for forensic science has been encouraged in the last years, motivated by the recently reopened debate about the infallibility of some classical forensic disciplines and the controversy about the admissibility of evidence in courts. Standardization of procedures, proficiency testing, transparency in the scientific evaluation of the evidence and testability of the system and protocols are emphasized in order to guarantee the scientific objectivity of the procedures. In this chapter those ideas and their relationship to automatic forensic speaker classification will be analyzed in order to define where automatic speaker classification is and which direction should it take under this context. Following the DNA methodology, which is being regarded as the scientific "golden" standard for evidence evaluation, the Bayesian approach has been proposed as a scientific and logical methodology. Likelihood ratios (LR) are computed based on the similarity-typicality pair, which facilitates the transparency in the process. The speaker classification is performed by the fact finder, who defines the possible hypotheses involved in the classification process. Thus, the prior probability of the hypotheses and the LR computed by the forensic system are used to assign a class to each suspected speaker depending on the defined hypotheses. The definition of this hypotheses typically refer to the speaker identity, thus leading to a speaker recognition task, but they can be defined in a more general context of speaker classification. The concept of calibration as a way of reporting reliable and accurate opinions is also addressed. Application-independent evaluation techniques (Cllr and APE curves) are addressed as a proper way for presenting results of proficiency testing in courts, as these evaluation metrics clearly show the influence of calibration errors in the accuracy of the inferential decision process. In order to illustrate the effects of calibration, we conclude with new experimental examples used as blind proficiency test following the NIST SRE 2006 evaluation protocol. [ABSTRACT FROM AUTHOR]

In this paper we study the application of user-dependent score fusion to multilevel speaker recognition. After reviewing related works in multimodal biometric authentication, a new score fusion technique is described. The method is based on a form of Bayesian adaptation to derive the personalized fusion functions from prior user-independent data. Experimental results are reported using the MIT Lincoln Laboratory's multilevel speaker verification system. It is experimentally shown that the proposed adapted fusion method outperforms both user independent and non-adapted user-dependent fusion approaches. [ABSTRACT FROM AUTHOR]

In this paper, the use of biometric systems in forensic applications is reviewed. Main differences between the aim of commercial biometric systems and forensic reporting are highlighted, showing that commercial biometric systems are not suited to directly report results to a court of law. We propose the use of a Bayesian approach for forensic reporting, in which the forensic scientist has to assess a meaningful value, in the form of a likelihood ratio (LR). This value assist the court in their decision making in a clear way, and can be computed using scores coming from any biometric system, with independence of the biometric discipline. LR computation in biometric systems is reviewed, and statistical assumptions regarding estimations involved in the process are addressed. The paper is focused in handling small sample size effects in such estimations, presenting novel experiments using a fingerprint and a voice biometric system. [ABSTRACT FROM AUTHOR]