Below is the abstract of the talk:
Voice recognition is the process by which distinctive characteristics of an individual’s speech are used to identify or verify who they are. As lay listeners, humans recognise familiar voices intuitively in an everyday sense and may also find themselves being ‘earwitnesses’ to a crime, albeit rarely. When carried out by trained practitioners using specialised methodologies and tools, voice recognition, comparing often unknown speech samples, can play an important role in investigative and forensic contexts.
This talk will consider the landscape of forensic voice recognition, encompassing auditory analysis by trained listeners, acoustic-phonetic measurements of perceptually salient features, and automatic speaker recognition using signal processing and modelling algorithms that are statistical or based on deep neural networks. The Bayesian likelihood ratio framework will be critically examined as a means of evaluating the strength of evidence using any voice analysis methodology. The importance of validation of the prevalent and emerging approaches, to understand their limitations and to provide reliable and transparent reports to the courts, will be discussed.
Additionally, the varying acceptance of voice recognition evidence in different parts of the world will be explored. Anticipating the new challenges posed by machine-created spoofed speech, this talk also will reflect on the risks, mitigations and, more optimistically, emerging opportunities afforded by using both human- and machine-based analysis.