SoniNet: an AI framework for enhancing voice features in speech-based health monitoring

Deep learning based speech embeddings learned from large-scale speech corpora have achieved strong performance across a range of speech analysis tasks. However, handcrafted acoustic features, such as prosodic, spectral, and voice-quality descriptors, remain widely used in clinical speech analysis because of their established role in characterizing vocal behavior. Despite this, these features often struggle to generalize across variations in speech dynamics, recording duration, and recording conditions. This paper introduces SoniNet , a framework that enhances the discriminative power of handcrafted acoustic features by learning data-driven temporal representations from large-scale speech data. SoniNet captures the temporal evolution of acoustic descriptors across speech recordings of variable length while maintaining compatibility with established acoustic feature sets used in clinical speech analysis. We evaluate SoniNet on benchmark speech data sets and demonstrate consistent improvements in classification performance across clinical speech assessment tasks, outperforming conventional acoustic modeling approaches in 22 out of 24 feature-task combinations. These findings highlight the potential of SoniNet as a robust framework for speech-based health monitoring, affect recognition, and computational behavioral analysis.