Implementation of forensic voice comparison within the new paradigm for the evaluation of forensic evidence

There is increasing pressure for forensic science to evaluate evidence in a logically correct manner, and to demonstrate validity and reliability under casework conditions (European Network of Forensic Science Institutes' 2015 Guideline for Evaluative Reporting, 2009 US National Research Council report, UK Forensic Science Regulator's 2014 Codes of Practice and Conduct). In forensic voice comparison the task of the forensic scientist is to assist the court to decide whether a recording of a voice of questioned identity (typically an offender) was produced by a speaker of known identity (typically a suspect) or not. The research presented here demonstrates the implementation of forensic voice comparison within a new paradigm for the evaluation of forensic evidence which consists of: 1. The use of the likelihood ratio framework, the logically correct framework for the evaluation of the strength of forensic evidence. 2. The use of relevant data (including data representative of the relevant population), quantitative measurements, and statistical models. 3. Empirical testing of validity and reliability of the forensic analysis system under conditions reflecting those of the case under investigation. 4. Transparency as to decisions made and procedures employed as part of the analysis. This dissertation illustrates the implementation of this paradigm via case studies based on three real legal cases. It describes solutions for dealing with challenges presented by real casework, including poor-quality recording conditions (such as combinations of reverberation, background noise, transmission-channel degradation, the speaker changing distance from the microphone, and very short recordings) and mismatches in recording conditions. Procedures include consideration of the appropriate defence hypothesis and hence the relevant population, selection of data representative of the relevant population, simulation of recording conditions reflecting those of the suspect and offender recordings, quantitative measurement and statistical modelling to calculate likelihood ratios, application of techniques to compensate for mismatch in recording conditions (including in the feature domain and in the i-vector domain), and empirical testing of the validity and reliability of the final system under conditions reflecting those of the case. The three case studies show how the general paradigm can be adapted to the challenges presented by different specific casework circumstances