Daniel Ramos-Castro

Robust estimation, interpretation and assessment of likelihood ratios in forensic speaker recognition

In this contribution, the Bayesian framework for interpretation of evidence when applied to forensic speaker recognition is introduced. Different aspects of the use of voice as evidence in the court are addressed, as well as the use by the forensic expert of the likelihood ratio as the right way to express the strength of the evidence. Details on computation procedures of likelihood ratios (LR) are given, along with the assessment tools and methods to validate the performance of these Bayesian forensic systems. However, due to the practical scarcity of suspect data and the mismatched conditions between traces and reference populations common in daily casework, significant errors appear in LR estimation if specific robust techniques are not applied. Original contributions for the robust estimation of likelihood ratios are fully described, including TDLRA (target dependent likelihood ratio alignment), oriented to guarantee the presumption of innocence of suspected but non-perpetrators speakers. These algorithms are assessed with extensive Switchboard experiments but moreover through blind LR-based submissions to both NFI-TNO 2003 Forensic SRE and NIST 2004 SRE, where the strength of the evidence was successfully provided for every questioned speech-suspect recording pair in the respective evaluations.

Likelihood Ratio Calibration in a Transparent and Testable Forensic Speaker Recognition Framework

A recently reopened debate about the infallibility of some classical forensic disciplines is leading to new requirements in forensic science. 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. Those ideas will be exploited in this paper in order to walk towards an appropriate framework for the use of forensic speaker recognition in courts. Evidence is interpreted using the Bayesian approach for the analysis of the evidence, as a scientific and logical methodology, in a two-stage approach based in the similarity-typicality pair, which facilitates the transparency in the process. The concept of calibration as a way of reporting reliable and accurate opinions is also deeply addressed, presenting experimental results which illustrate its effects. The testability of the system is then accomplished by the use of the NIST SRE 2005 evaluation protocol. Recently proposed application-independent evaluation techniques (Cllr and APE curves) are finally 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

Speaker verification using speaker- and test-dependent fast score normalization

A novel score normalization scheme for speaker verification is presented. The proposed technique is based on the widely used test-normalization method (Tnorm), which compensates test-dependent variability using a fixed cohort of impostors. The new procedure selects a speaker-dependent subset of impostor models from the fixed cohort using a distance-based criterion. Selection of the sub-cohort is made using a distance measure based on a fast approximation of the Kullback-Leibler (KL) divergence for Gaussian mixture models (GMM). The proposed technique has been called KL-Tnorm, and outperforms Tnorm in computational efficiency. Experimental results using NIST 2005 Speaker Recognition Evaluation protocol also show a stable performance improvement of our method on standard speaker recognition systems.

Suspect-Adapted MAP Estimation of Within-Source Distributions in Generative Likelihood Ratio Estimation

In this paper, a novel suspect-adaptive technique for robust Bayesian forensic speaker recognition via maximum a posteriori (MAP) estimation is presented, which addresses likelihood ratio (LR) computation in limited suspect speech data conditions obtaining good calibration performance. Robustness is achieved by the use of speaker-independent information, adapting it to the specificities of the suspect involved in the process. Thus, this procedure allows the system to weight the relevance of the suspect specificities depending on the amount of suspect data available via MAP estimation. Experimental results show robustness to suspect data scarcity and stable performance for any amount of suspect material. Also, the proposed technique outperforms other previously proposed non-adaptive approaches. Results are presented as discrimination capabilities (DET plots), distributions of LRs (Tippett plots) and expected cost of wrong decisions over any prior or decision cost (Cllr). The use of such evaluation metrics allows us to highlight the importance of LR calibration in the performance of a forensic system

Between-Source modelling for likelihood ratio computation in forensic biometric recognition

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.

Speaker Recognition The A TVS-UAM System at NIST SRE 05

Automatic speaker recognition systems have been largely dominated by acoustic-spectral-based systems, relying in proper modelling of the short-term vocal tract of speakers. However, there is scientific and intuitive evidence that speaker-specific information is embedded in the speech signal in multiple short- and long-term characteristics. In this work, a multilevel speaker recognition system combining acoustic, phonotactic, and prosodic subsystems is presented and assessed by blind submission to NIST 2005 Speaker Recognition Evaluation