A1934
Title: Advancing psychosocial health assessment with Automatic Speech Analytics Program (ASAP)
Authors: Amanda Chu - The Education University of Hong Kong (China) [presenting]
Abstract: ASAP is developed as a machine learning-based system designed for early detection of psychosocial distress, with a focus on statistical methods for linguistic and acoustic feature extraction. ASAP records client Speech, transcribes conversations in real time, delivers immediate risk evaluations through an intuitive interface, and generates automated reports to reduce administrative workload. Linguistic features were obtained by transcribing 100 audio recordings and developing a project-specific dictionary of 53 words, organized into 14 topics and four themes. Each word was mapped to three risk indicators: the Beck Depression Inventory-II, Caregiver Burden Inventory, and Family Resilience Assessment Scale. Principal component analysis was applied to reduce dimensionality and minimize redundancy and bias. Acoustic features were extracted using openSMILE, encompassing 65 low-level descriptors related to pitch, voice quality, spectral, and energy characteristics, which were further expanded into 6,503 statistical features representing waveform variability. The integration of these linguistic and acoustic features enabled accurate psychosocial risk classification, achieving 86\% accuracy in differentiating stress levels. The combination of dictionary-based linguistic analysis with advanced signal processing and statistical techniques highlights the potential of feature engineering in Speech Analytics for improving Resilience Assessment and Depression detection.
