Considering Accent Recognition Technology for Forensic Applications

Speaker recognition technology is becoming more available to forensic speech analysts to help to arrive at conclusions around how likely the speech in multiple recordings was produced by the same speaker. However, there is not currently a suitable technological tool that could assist with speaker profiling tasks (i.e. tasks where we wish to deduce information about an unknown speaker). Accent recognition technology could play a role in speaker profiling tasks. This thesis therefore presents numerous automatic accent recognition experiments that have been motivated by forensic applications.

This thesis conducts a detailed examination of one automatic accent recognition system in particular, the York ACCDIST-based automatic accent recognition system (the Y-ACCDIST system). It is trained to assign an accent label to a speaker's speech sample. Unlike other accent recognition system architectures, Y-ACCDIST takes a segmental approach by forming models of speakers' accents using representations of individual phonemes. Implementing a segmentation phase comes at a practical cost, but it is expected that Y-ACCDIST's segmental approach captures a more detailed reflection of a speaker's accent than other accent recognition systems. When classifying speech samples into one of four categories, Y-ACCDIST achieved a recognition rate of 86.7% correct, while the best-performing text-independent system obtained 47.5%.

This thesis also shows Y-ACCDIST's performance on spontaneous speech data. On a three-way classification task on Northern English accents, we witness a recognition rate of 86.7% correct. Additionally, we achieved 63.1% correct when classifying recordings into one of seven non-native English categories. The latter task is also a demonstration of Y-ACCDIST's capabilities on telephone data.