Towards improving transcripts of audio recordings in the criminal justice system

Transcripts of speech evidence are often presented to juries for the purpose of providing a written record of speech within an audio recording and/or aiding members of the court in understanding the speech content of poor quality recordings. The accuracy and impartiality of these transcripts is paramount, but such qualities require systematic and scientifically-informed methods. There are three main aims of this thesis: (1) to gain a better understanding of current practices when transcribing poor quality evidential audio, (2) to explore how background noise and regional accents affect the content of transcripts, and (3) to develop a method for evaluating transcripts for forensic purposes. The first aim is explored through a survey of expert transcription practices and a focus interview regarding common issues encountered in non-expert transcripts; both studies demonstrate variability in the methods employed to produce transcripts. A push towards standardisation is recommended, encouraging (a) further research on transcription methods, focusing on method validation and proficiency testing, and (b) the production of standards and/or guidelines. The second and third aims of the thesis are addressed in two studies in which human and automatic transcription performance is compared across different audio qualities and accents. Findings reveal that transcripts are significantly worse for audio with increased background noise and substantially worse for unfamiliar accents. A new forensically-motivated method of evaluating transcripts is employed in these studies, focusing on substitution errors and their potential impact on meaning; this can be used (or developed) for further research and proficiency testing. The work in this thesis shows a huge research gap concerning the production of transcripts for use in the criminal justice system, which needs to be addressed by further empirical testing of transcription methods, human and automatic performance, and human-automatic hybrid approaches to transcription.