Forensic phonetics and automatic speaker recognition

Abstract

Speech technologists are making rapid advances in terms of the performance of Automatic Speaker Recognition (ASR) systems, which are designed to carry out spectral comparisons of speech recordings that are of evidential significance in criminal investigations. ASR systems yield quick, inexpensive and scientifically valid results, which the authorities in some jurisdictions see as superior to those produced by human speech experts. Many governments are therefore investing heavily in training police officers and forensic laboratory staff in the use of commercially available ASR products. However, there are aspects of human psychology and behaviour which influence the acoustic-phonetic characteristics of speech but which have hitherto attracted relatively little forensically oriented empirical research. Since forensic casework recordings so often contain speech produced by individuals under high levels of emotional arousal and/or psychological stress, it is vital – not least if we wish to keep improving ASR system performance – that we better understand the effects that these non-linguistic factors can have on the speech signal. We argue that the introduction of ASR will not necessarily replace human expert speech analysts. Instead, it is opening up new kinds of challenge for forensic speech scientists, forensic linguists and forensic psychologists, and encouraging collaborations based on their complementary expertise.