Lightweight Edge-Based Speech Emotion Recognition with Human-Centered and Bio-Inspired Design for Early Mental-Health Detection

Mental-health disorders such as depression, anxiety, and chronic stress remain widespread, while early assessment is often hindered by stigma, limited resources, and the lack of continuous monitoring tools. This paper presents a lightweight Speech Emotion Recognition (SER) system designed for embedded edge devices to support real-time, privacy-preserving emotion assessment. The proposed framework applies standard preprocessing, including silence removal and voice activity detection, and extracts complementary Mel-Frequency Cepstral Coefficients (MFCC), Fractional Frequency Cepstral Coefficients (FrFCC), and prosodic features. A bio-inspired attention mechanism adaptively fuses these feature streams before classification using a depthwise-separable convolutional neural network optimized through pruning and quantization for TensorFlow Lite deployment. Experiments on the EMO-DB, RAVDESS, and TESS datasets achieve accuracy above 90% and unweighted average recall above 88%. Deployment on Raspberry Pi 4 and a mid-range Android device demonstrates inference latency below 150 ms and end-to-end delay under 500 ms with low power consumption, enabling continuous ondevice operation. Ablation studies confirm the contribution of each feature set and the effectiveness of attention-based fusion. Preliminary user evaluations indicate positive usability, supporting the feasibility of the proposed system as an ondevice tool for mental-health monitoring support.