Search

Imagine you are a police officer investigating a crime in which somebody has made a threatening phone call. There may not be any physical evidence like DNA or fingerprints, but the person’s voice is captured on a recording. Experts can analyze the voice recording and compare it with the voice of a known suspect to find out how likely it is that they are the same person. This method is called forensic voice comparison. It works because people’s voices contain lots of information about them. Your voice tells a story about where you grew up and learned to speak, and your voice also depends on your individual biological make-up. In this article, we explain how this information is used to compare recordings of voices in criminal cases.

Since the 1960s, there have been calls for forensic voice comparison to be empirically validated under casework conditions. Since around 2000, there have been an increasing number of researchers and practitioners who conduct forensic-voice-comparison research and casework within the likelihood-ratio framework. In recent years, this community of researchers and practitioners has made substantial progress toward validation under casework conditions becoming a standard part of practice: Procedures for conducting validation have been developed, along with graphics and metrics for representing the results, and an increasing number of papers are being published that include empirical validation of forensic-voice-comparison systems under conditions reflecting casework conditions. An outstanding question, however, is: In the context of a case, given the results of an empirical validation of a forensic-voice-comparison system, how can one decide whether the system is good enough for its output to be used in court? This paper provides a statement of consensus developed in response to this question. Contributors included individuals who had knowledge and experience of validating forensic-voice-comparison systems in research and/or casework contexts, and individuals who had actually presented validation results to courts. They also included individuals who could bring a legal perspective on these matters, and individuals with knowledge and experience of validation in forensic science more broadly. We provide recommendations on what practitioners should do when conducting evaluations and validations, and what they should present to the court. Although our focus is explicitly on forensic voice comparison, we hope that this contribution will be of interest to an audience concerned with validation in forensic science more broadly. Although not written specifically for a legal audience, we hope that this contribution will still be of interest to lawyers.

Among phoneticians, the Vocal Profile Analysis (VPA) is one of the most widely used methods for the componential assessment of voice quality. Whether the ultimate goal of the VPA evaluation is the comparative description of languages or the characterization of an individual speaker, the VPA protocol shows great potential for different research areas of speech communication. However, its use is not without practical difficulties. Despite these, methodological studies aimed at explaining where, when and why issues arise during the perceptual assessment process are rare. In this paper we describe the methodological stages through which three analysts evaluated the voices of 99 Standard Southern British English male speakers, rated their voices using the VPA scheme, discussed inter-rater disagreements, and eventually produced an agreed version of VPA scores. These scores were then used to assess correlations between settings. We show that it is possible to reach a good degree of inter-rater agreement, provided that several calibration and training sessions are conducted. We further conclude that the perceptual assessment of voice quality using the VPA scheme is an essential tool in fields such as forensic phonetics but, foremost, that it can be adapted and modified to a range of research areas, and not necessarily limited to the evaluation of pathological voices in clinical settings.

This paper contributes to the ongoing discussion about the distinction between observations and propositions in forensic inference, with a specific focus on forensic voice comparison casework conducted in the UK. We outline both linguistic and legal issues which make the evaluation of voice evidence and the refinement of propositions problematic in practice, and illustrate these using case examples. We will argue that group-level observations from the offender sample will always be evidential and that the value of this evidence must be determined by the expert. As such, a proposal is made that experts should, at least conceptually, think of voice evidence as having two levels, both with evidential value: group-level and individual-level. The two rely on different underlying assumptions, and the group-level observations can be used to inform the individual-level propositions. However, for the sake of interpretability, it is probably preferable to present only one combined conclusion to the end user. We also wish to reiterate points made in previous work: in providing conclusions, the forensic expert must acknowledge that the value of the evidence is dependent on a number of assumptions (propositions and background information) and these assumptions must be made clear and explicit to the user.

In likelihood ratio (LR)-based forensic speaker\ud comparison it is essential to consider correlations between parameters to accurately estimate the overall strength of the evidence. Current approaches attempt to deal with correlations after the computation of LRs (back-end processing). This paper explores alternative, front-end techniques, which consider the underlying correlation structure of the raw data. Calibrated LRs were computed for a range of parameters commonly analysed in speaker comparisons. LRs were combined using (1) an\ud assumption of independence, (2) the mean, (3)\ud assumptions from phonetic theory, and (4) empirical correlations in the raw data. System (1), based on an assumption of independence, produced the best validity (Cllr = 0.04). Predictably, overall strength of evidence was also highest for system (1), while strength of evidence was weakest using the mean (2). Both systems (3) and (4) performed well achieving Cllr values of ca. 0.09.

Within the field of forensic speech science there is increasing acceptance of the likelihood ratio (LR) as the logically and legally correct framework for evaluating forensic voice comparison (FVC) evidence. However, only a small proportion of experts cur- rently use the numerical LR in casework. This is due primarily to the difficulties involved in accounting for the inherent, and arguably unique, complexity of speech in a fully data-driven, numerical LR analysis. This thesis addresses two such issues: the definition of the relevant population and the amount of data required for system testing. Firstly, experiments are presented which explore the extent to which LRs are affected by different definitions of the relevant population with regard to sources of systematic sociolinguistic between-speaker variation (regional background, socio-economic class and age) using both linguistic-phonetic and ASR variables. Results show that different definitions of the relevant population can have a substantial effect on the magnitude of LRs, depending on the input variable. However, system validity results suggest that narrow controls over sociolinguistic sources of variation should be preferred to general controls. Secondly, experiments are presented which evaluate the effects of development, test and reference sample size on LRs. Consistent with general principles in statistics, more precise results are found using more data across all experiments. There is also considerable evidence of a relationship between sample size sensitivity and the dimensionality and speaker discriminatory power of the input variable. Further, there are potential trade-offs in the size of each set depending on which element of LR output the analyst is interested in. The results in this thesis will contribute towards im- proving the extent to which LR methods account for the linguistic-phonetic complexity of speech evidence. In accounting for this complexity, this work will also increase the practical viability of applying the numerical LR to FVC casework.

The likelihood ratio (LR) is the “logically and legally correct” (Rose and Morrison 2009:143) framework for the estimation of strength-of-evidence under two competing hypotheses. In forensic voice comparison these considerations are reduced to the similarity and typicality of features across a pair of suspect and offender samples. However, typicality can only be judged against patterns in the relevant population (Aitken and Taroni 2004:206). In calculating numerical LRs typicality is assessed relative to a sub-section of that population. This study considers issues of variability relating to the delimitation of reference data with regard to the number of speakers and number of tokens per speaker. Using polynomial estimations of F1 and F2 trajectories from spontaneous GOOSE (Wells 1982), LR comparisons were performed against a reference set of up to 120 speakers and up to 13 tokens per speaker. Results suggest that mean same-speaker LRs are robust to such variation until the reference data is limited to small numbers of speakers and tokens. However, variance and severity of error may be continually reduced with the inclusion of more data. The definition of the relevant population with regard to regional variety is also assessed. Results for LRs are presented across four sets of test data where only one set matches the reference population for accent. In the absence of differences in levels of within-speaker variation, the magnitude of same-speaker LRs and severity of error are shown to be considerably higher for the ‘mismatch’ test sets. However, results indicate that the removal of regionally-defining acoustic information may reduce the effect of accent divergence between the evidential and reference data. This has positive implications for the application of the numerical LR approach.