Holistic Affect Recognition Using PaNDA: Paralinguistic Non-Metric Dimensional Analysis

Humans perceive emotion from each other using a holistic perspective, accounting for diverse personal, non-emotional variables, such as age and personality, that shape expression. In contrast, today’s algorithms are mainly designed to recognize in isolation, usually demonstrated only within one relatively narrow database. this article, we propose multi-task learning approach jointly learn the recognition of affective states speech along with various speaker attributes. A problem is sometimes inductive transfer can negatively impact performance. To mitigate negative transfer, introduce Paralinguistic Non-metric Dimensional Analysis (PaNDA) method systematically measures task relatedness also enables visualizing topology phenomena whole. addition, present generic framework conflates concepts single-task learning. Using framework, construct two models demonstrate affect recognition: treats all tasks equally related, whereas incorporates correlations between main its supporting obtained PaNDA. Both employ deep neural network, which separate output layers used predict discrete continuous attributes, while hidden shared across different tasks. On average 18 classification regression tasks, weighted PaNDA significantly improves performance compared unweighted