Towards improving transcripts of audio recordings in the criminal justice system

Evidence presented in legal cases often features recordings of speech; perhaps the suspect allegedly confessed to the crime in a police interview, or CCTV footage captured the audio from an off-camera altercation at the crime scene. The evidential audio recording is often accompanied by a transcript, which is a written record of the speech content. While the transcript itself is not the evidence, it is very heavily relied upon and commonly used in place of the audio recording. However, the complexity of transcription is often underestimated: it is not as simple as ‘writing down what someone says’. There are subconscious processes, as well as many conscious decisions, that take place during transcription. These can have a substantial impact on what is transcribed and how that content is later interpreted. Given that evidential transcripts can contribute to justice being served, it is crucial that they are produced using transparent, thoroughly-investigated methods. However, there is currently a major research gap concerning forensic transcription, with little to no work exploring how these transcripts are or should be produced. My PhD research investigated the methods that are currently employed to produce transcripts for use within the criminal justice system, and explored ways to improve these methods and therefore the quality and impartiality of transcripts. Findings showed that experts use carefully-managed processes to produce transcripts, but there is variability in their practices due to a lack of research and guidelines. During the fellowship, I will expand the small body of work on forensic transcription through the publication of multiple academic journal articles and presentations at international conferences. Attendance at world-leading conferences will publicise my research and allow me to network with researchers across multiple fields of linguistics and speech technology. I will also set up the International Network for Forensic Transcription, which will bring together academic researchers and industry professionals. The aims of the network are to (i) bridge the gap between academia and industry, ensuring that research aims are driven by the needs of those conducting casework, (ii) provide a space where research findings can be shared and discussed, and (iii) take significant steps towards guidelines or a framework for practice to address regulatory needs. Furthermore, non-expert (e.g. police) transcribers approach transcription in a more ad hoc manner, which is evidenced by a lack of consistency, structure and guidance. In my PhD, I conducted the first study into using AI-based automatic speech recognition (ASR) systems to produce verbatim transcripts of speech in the context of police-suspect interviews. It is easy to see the appeal of such technology, but it is crucial that the dangers are also considered. My research findings showed that transcripts produced by commercial ASR systems will always require checking and correcting by a human transcriber to produce a reliable and accurate record of the speech. This is a crucial step in ensuring that good quality transcripts make their way to juries. During the fellowship, I will develop and submit a funding bid for a large research project investigating the use of speech technologies in the transcription of police interviews. The output of the project will be a pipeline for the safe integration of ASR into policing practices, which can be disseminated to UK police forces through training courses run by the Forensic Speech Services unit at the University of York.