Forensic speaker identification

Project within the UROP scheme, initiated by Kirsty McDougall, work carried out by Qinkan Wu

A growing number of court cases involve the need to establish the speaker of some recorded speech - a hoax emergency call, a fraudulent phone transaction, an obscene voicemail, the planning of a drug deal and so on. Voices, however, are not like fingerprints. Contrary to the impression given by television shows such as CSI, there is no technique for identifying a speaker with 100% reliability. A person’s voice varies, depending on tiredness, emotion, how loud and fast he or she is speaking, and many other factors. This variability within an individual’s voice makes the task of speaker identification very complicated in the forensic domain, where recordings are usually short, of poor quality and in a range of speaking styles.

The DyViS project (‘Dynamic Variability in Speech: A Forensic Phonetic Study of British English’) in the Phonetics Laboratory in the Department of Linguistics is carrying out research tackling these issues. The project has compiled a large-scale database of speech in different speaking styles from 100 speakers of Standard Southern British English. These recordings are being analysed to determine how well the speakers can be discriminated and what the best measures are for characterising their speech. The effects of using the telephone on an individual's speech is also being analysed.

The student project will involve developing scripts to process and analyse speech files efficiently using programs such as Praat. A background in computer programming is required, but experience in using Praat is not necessary. The exact nature of the work to be undertaken is flexible depending on the skills and interests of the student.