Speech Recognition in Policing

Producing a written record of a conversation provides an accessible way for information to be reviewed. Speech differs from writing in many ways, especially when compared with standard written language. However, a written document can sometimes be a practical and efficient way for users to access information. This is often the case for police interviews, these are usually long and only certain parts of information need to be accessed by a given individual at any one point. Transcription of police interviews by humans is a labour intensive process, with police forces dedicating significant resources to facilitate it. Thousands of documents are produced annually and support countless investigations. However, there is currently no standardised framework; each of the 43 police forces in England and Wales has its own internal process. The lack of consistency is reflected in the type of transcript produced: for example, some produce verbatim (word-for-word) transcripts, whereas other forces produce only summaries of interviews. 

AI-based automatic speech recognition (ASR) could facilitate standardisation and save time. A well designed and safely integrated tool could provide a standardised framework for all forces, speed up transcription, effectively facilitate human checking and potentially support playback in court (provided there is an appropriate technological set up). However, there are increasing examples of the risks involved with poor deployment of AI-based ASR systems in the public sector.