This thesis introduces and explores the performance of the Y-ACCDIST system (the York ACCDIST automatic accent recognition system). Based on the ACCDIST metric (Huckvale, 2004), it is a newly developed accent recognition system intended for forensic applications. Accent has received a lot of research attention within speech technology as it is often to blame for automatic speech recognition errors. A lot of research has therefore targeted automatic accent recognition while taking the automatic speech recognition application into account. Little has been done, however, to research automatically recognising speakers’ accents for forensic purposes. Such a task might involve identifying speaker properties (e.g. geographical origin) if no suspect is in the frame for making an incriminating telephone call.

The Y-ACCDIST system is applied to the forensic context in two main ways. Firstly, it is applied to geographically-proximate accents, where a predicted increase in similarity between varieties exists. This accent recognition task is therefore expected to be more difficult than tasks in previous studies. Secondly, the model is adapted in such a way which makes it possible to process spontaneous speech, instead of just highly comparable speech content (i.e. read prompts). The present these shows accent recognition results, distinguishing between four geographically-proximate accents, of up to 90%. Accent recognition results of spontaneous speech are lower (up to 59.3%). However, light is shed on clear research directions aiming to improve this result.