Analysing the performance of automated transcription tools for covert audio recordings

The orthographic transcription of audio recordings can provide important evidence in a forensic case (Fraser, 2021), but producing transcripts is an extremely time-consuming task and is often a prerequisite to further analyses. • Huge improvements in automatic speech recognition (ASR) have been observed throughout the past two decades, particularly with the recent development of deep learning (Xiong et al., 2016). • The use of ASR could significantly decrease the amount of time and effort taken to produce a transcript and this could make such systems an attractive prospect to those in law enforcement (Loakes, 2022). • The appropriacy of ASR for the transcription of indistinct forensic-like audio is worthy of investigation. This paper reports the design and results of a controlled transcription experiment in which twelve automated transcription tools produced transcripts for the same audio recording.