The Laboratory conducts research and development aimed at improving casework capabilities in forensic speech science, particularly casework conducted within the new paradigm for the evaluation of forensic evidence, i.e.:
- Quantification of strength of evidence as likelihood ratios (the likelihood-ratio framework is the logically correct framework for evaluation of forensic evidence)
- Likelihood ratios calculated using relevant data, quantitative measurement, and statistical models
- Validation of system performance under conditions reflecting those of the case under investigation
- Reduction of the potential for cognitive bias by restricting subjective judgements to matters such as selection of appropriate data to enter into the system and by directly reporting the output of the statistical model as the strength of evidence
The primary focus of the Laboratory is on forensic voice comparison (aka forensic speaker recognition), but Laboratory members have also worked on disputed-utterance analysis and on speaker recognition by earwitnesses.
FSSL works closely with the Forensic Data Science Laboratory.
- Directors, Academic and Research Staff
Dr Nabanita Basu
Research Associate in Forensic Data Science,
Dr Philip Weber
Research Fellow in Forensic Data Science
Dr Geoffrey Stewart Morrison
Director of the Forensic Speech Science Laboratory
- Honorary and Adjunct Members
Dr Ewald Enzinger
Visiting Research Fellow
Senior Research Engineer, Eduworks Corporation
Prof Cuiling Zhang
Visiting Research Fellow
Chongquing Institutes of Higher Education Key Forensic Science Laboratory
Dr Claudia Rosas
Visiting Research Fellow
Associate Professor, Instituto de Lingüística y Literatura, Universidad Austral de Chile
- Development of a forensic voice comparison system
We are developing a forensic voice comparison system that can be used for research and casework. We view a system for conducting forensic voice comparison as not simply a collection of tools, but also protocols, databases suitable for training and testing under casework conditions, documentation, validation reports, and well trained practitioners. We aim to develop a system that will meet legal admissibility requirements such as those of Federal Rule of Evidence 702 and the Daubert trilogy of Supreme Court rulings in the United States, and of Criminal Practice Directions 19A in England & Wales.
In forensic voice comparison casework, the relevant population and the recording conditions vary greatly from case to case. Researchers and practitioners need protocols, tools, and data that provide them with the flexibility to deal with this case to case variability. Practitioners need to be able to train (or retrain or optimize) the system for the conditions of the case, and they need to be able to empirically validate the performance of the system under conditions reflecting those of the case. In order to inform practice, researchers need to explore which options and settings give best performance under particular conditions, and explore the robustness of systems to variability in conditions.
Commercially marketed software tools often lack flexibility, and may be too expensive for researchers and practitioners in lower GDP countries. Many researchers and practitioners in the field lack the programming skills to make use of existing open-source automatic speaker recognition toolsets, and licencing restrictions may prevent such toolsets from being used for casework which counts as commercial activity. For different reasons, existing commercial and open-source tools are often insufficiently well documented for end users and others to be able to easily understand what the tools are actually doing. This is in opposition to the transparency that may be required by the courts. Researchers and practitioners therefore need software tools that are low cost, flexible, and easy to use (controllable via GUI or only requiring very limited programming skills), that are very well documented, that are designed to facilitate validation, and that have code that is open to inspection (we envisage open source, but not open distribution).
In the context of this research and development project, we define two groups of end-users: 1. Researchers and practitioners who will (potentially) use the system to do research and to conduct casework. 2. Service users, i.e., organizations that commission practitioners to perform forensic voice comparison analyses. Potential service users include defence lawyers and law-enforcement agencies.
We have established relationships with several collaborators. Collaboration ranges from service users helping with end-user needs assessment, through to researchers collecting and sharing data and actively contributing to research and development.
Universidad Austral de Chile – Claudia Rosas, Jorge Sommerhoff
Policia de Investigaciones
Southwest University of Political Science and Law – Cuiling Zhang
German Federal Police, Bundeskriminalamt (BKA) – Michael Jessen
Netherlands Forensic Institute (NFI) – David van der Vloed
- United Kingdom:
National Crime Agency (NCA)
- United States of America:
Federal Bureau of Investigation (FBI) – David Marks
- 2018–2019 we worked on end-user needs assessments to identify the data that need to be collected, and the tools, protocols, and training programmes that need to be developed.
- 2018–2019 we began development of prototypes for core software tools. We will be developing a system that uses x-vectors.
- Research England: Expanding Excellence in England (E3) fund.