Seminar | Exploring individual-level variability in automatic speaker recognition performance
Seminars / Lectures / Workshops
-
Date
23 Sep 2026
-
Organiser
Department of English and Communication
-
Time
17:00 - 18:00
-
Venue
Online via Zoom
Speaker
Prof. Vincent Hughes
Summary
In this talk, I will cover outcomes from the Person-specific Automatic Speaker Recognition (PASR) project led by the University of York, together with Oxford Wave Research and the Netherlands Forensic Institute. The talk will be in two parts. The first will explore automatic speaker recognition performance with a small, controlled set of data collected at York called SISTEM (Speech of Individual Speakers Across Time by Experts in Multiple Sessions). The corpus allows us to systematically assess when and how systems perform well and where they struggle. I will also discuss testing conducted with a very large, forensically realistic corpus provided by the UK Government. Our testing goes beyond the typical high-level metrics of overall system performance, to assess the behaviour of individual speakers and individual files within our dataset. The aim is to describe the range of speaker and file variability and then attempt to explain and even predict speakers and files that may produce poor performance. I will also report on testing conducted to assess the stability of individual speakers across generations of automatic speaker recognition systems.
Keynote Speaker
Prof. Vincent Hughes
Professor, University of York, United Kingdom
Vincent Hughes is Professor of Forensic Speech Science in the Department of Language and Linguistic Science at the University of York. A central theme of his research involves assessing how individual voices differ from each other. This involves testing and validating phonetic and AI-based methods for forensic voice comparison and the statistical evaluation of expert forensic evidence. He has recently completed a large-scale ESRC-funded projects looking at the use of automatic speaker recognition systems as forensic evidence and has also led interdisciplinary projects focused on the individuality of the human voice. More recently, he has worked on applying forensic approaches to the analysis of deepfake audio and in the context of transcription in law enforcement. Professor Hughes is also co-lead of Forensic Speech Services at York, which bridges the gap between research and practice through forensic casework, CPD and training courses, consultancy services, and collaborative and contract research.
