Dynamic minimum pause threshold estimation for speech analysis in studies of cognitive function in ageing

Cognitive decline represents the biggest limiting factor to independence in older adults. Speech analysis has emerged as an alternative to standard cognitive assessment tools. Temporal segmentation of speech is reported in many studies and typically employs a static threshold to define a pause. This study investigated the effect of using pause and utterance duration distribution data in differentiating between cognitively healthy and impaired older adults. Three sets of features were extracted from 187 speech recordings: temporal features using a static 250ms threshold; temporal features using a dynamic threshold; and pause and utterance duration distribution parameters. The ability of each of these sets to differentiate between cognitively healthy and cognitively impaired participants was investigated using a Linear Discriminant Analysis (LDA) classifier. Improvements of 0.22% (to 64.20%) in sensitivity, 6.33% (73.12%) in specificity, and 3.27% (68.66%) in overall accuracy were observed in the performance of the classifier using the pause and utterance duration distribution parameters when compared to the static temporal features. The use of the dynamic threshold had a negative impact on the classifier performance, with a decrease of 5.73% (to 58.25%) in sensitivity, 1.10% (65.69%) in specificity, and 3.42% (61.97%) in accuracy.