The effects of speaker individuality on voice disguise and speaker recognition in application to forensic speech science

The human voice is widely viewed as an important marker of identity, anchoring an individual to their body. The hypothesised uniqueness of individual voices lies at the foundation of research in forensic speech science, which is an area of applied speech science dedicated to the use of voice evidence in criminal investigations. Typical forensic casework involves establishing whether or not two voice recordings are likely to be from the same speaker, or whether they could have been produced by different speakers (Nolan, 1991). One area that provides a significant challenge to this task is vocal disguise, whereby an individual has purposefully modified their voice in order to sound like somebody else and, therefore, evade identification. This is particularly relevant in the context of
3
detecting instances of identity fraud (especially in the current climate of voice identification technologies implemented by banks and other large organisations), with an increase of 15% in the past year of overall fraud cases (Office of National Statistics, 2019). Worryingly, past research has found that some individuals can imitate the speech of others to a high degree of accuracy (Leemann & Kolly, 2015). Despite this, research to date has not comprehensively addressed the precise mechanisms of voice disguise and whether even the best disguises contain features that could identify the speaker. As a consequence, it can be extremely difficult to provide a robust conclusion on whether a disguised voice is likely to be the suspect or not, as specific evidence from the speech signal must be presented in order to justify such claims in court.