Investigating the phonetic properties of discourse-pragmatic markers with implications for forensic voice comparison

Features of the voice which have high between-speaker variation and low within-speaker variation are advantageous for forensic voice comparison (FVC). Analysis often involves segmental and suprasegmental variables. Phonetic variation at the word level, however, is something that has been little explored. As there is no guarantee that multiple forensic recordings will share the same features, the most frequently-used words in speech may be of value to the FVC practitioner. In order to explore this, three of the most frequent words in British English (like, yeah, just) were selected for analysis, referred to as discourse-pragmatic markers (DPMs). The restrictions on DPM phonetic variation generally are under-researched. Theories of grammaticalisation and usage-based phonology make various predictions around how a frequent DPM will be pronounced, but there are few studies which assess these together. This thesis therefore has two interwoven aims: (1) assess the effects of pragmatic function and frequency of use in context on the phonetic variation of DPMs. (2) assess the strength of forensic voice comparison evidence based on the phonetic variation of DPMs. (1) is explored through three studies which show that DPMs vary phonetically in significant ways according to their function; when DPMs are used in newer ways (e.g. discourse marking or expressing epistemic uncertainty) they show consistently higher degrees of phonetic reduction than when they are used in canonical or older ways (e.g. as adverbs or conjunctions). This suggests that the pronunciation differences embodied by DPMs reflect a separation of one item into multiple lexical items. (2) is explored through three studies which assess the speaker discriminant capability of DPMs against already-established FVC features. These studies found that just performed well as a speaker discriminant, but yeah and like had a high level of within-speaker variation. This can be explained partly by looking at the patterns found to affect DPM phonetic realisation. Surrounding contexts and pragmatic functions affected the entire vocalic sequence in yeah and like, whereas the vowel quality of just was more stable due to its position between obstruents. These findings show that speakers utilise phonetic resources to indicate discourse-pragmatic functions, and therefore caution against utilising DPMs in forensic voice comparison testing when there is a lack of functional and contextual equivalence between tokens.