Speech Analytics for the Detection of Neurological Conditions in Global Varieties of English

Dementia refers to the decline of cognition and memory due to neurological conditions. Both speech content and acoustics can be analysed as biomarkers to detect dementia, offering a promising, cost-effective alternative to traditional methods. While speech-based dementia classification has shown success for L1 (native) speakers, this thesis explores the challenges posed by L2 (non-native) English speakers.

Investigating Automatic Speech Recognition (ASR) and its performance across global English varieties, we identify a disconnect between advertised and real-world system capabilities. ASR systems often reduce L2 language diversity, treating fine-tuned derivations of the same models as unique systems. We analyse ASR evaluation methods and propose a new multi-metric standard inspired by advances in Machine Translation. Examining ASR performance in dementia datasets, we find only a small deviation between L1 and L2 speech using the latest models, suggesting potential reliability for transcribing non-native speech. However, L2 speech produces speech analytics that cluster more closely with L1 speakers with dementia than with healthy L1 speakers, particularly in lexical and syntactic complexity metrics. This thesis proposes a framework for classification pipeline evaluation focused on measuring feature fluctuation as an interpretable framework for understanding inaccuracies within analytics directly fed into our classification models.

This system design approach aims to support clinician-facing reports and improve awareness of healthcare equity outcomes. To deploy these systems, we must better understand how different configurations impact minority voices. We propose longitudinal, speaker-dependent analytics as a means of calibrating dementia classification pipelines to a profile of speech analytics rather than single-session data. This approach would contextualise ‘healthy’ in the context of each speaker’s voice. Additionally, we suggest redefining healthy control classifications to better reflect cognitive concerns. Finally, we advocate for a more integrated approach to Speech Technology development for dementia detection, ensuring system outputs align with clinician-facing reports, patient-facing reports, and assurance protocols.