Incorporating automatic speech recognition methods into the transcription of police-suspect interviews: factors affecting automatic performance

Introduction In England and Wales, transcripts of police-suspect interviews are often admitted as evidence in courts law. Orthographic transcription is a time-consuming process usually carried out by untrained transcribers, resulting records that contain summaries large sections the interview paraphrased speech. The omission or inaccurate representation important speech content could have serious consequences court It therefore clear investigation into better solutions for police-interview required. This paper explores possibility incorporating automatic recognition (ASR) methods process, with goal producing verbatim without sacrificing police time money. We consider potential viability “first” draft would be manually corrected transcribers. study additionally investigates effects audio quality, regional accent, ASR system used, well types magnitude errors produced their implications context transcripts. Methods Speech data was extracted from two forensically-relevant corpora, speakers accents British English: Standard Southern English West Yorkshire (a non-standard variety). Both high quality degraded version each file transcribed using three commercially available systems: Amazon, Google, Rev. Results System performance varied depending on while accent not found to significantly predict word error rate, distribution substantially across accents, more potentially damaging English. Discussion low rates easily identifiable Amazon suggest incorporation viable, though work required investigate other contextual factors, such multiple different background noise.