Speech Stress Detection In Marathi And Susas Databases Using Weight-optimized Neural Networks

Abstract

Stress Profoundly Alters Human Cognitive And Physiological States, Making Early And Automated Detection A Critical Technological Goal [1]. This Study Presents A Streamlined Speech Emotion Recognition (SER) System Engineered For Accurate Stress Classification [2]. The Methodology Operates Across Two Major Domains: A Manual Feature Architecture Integrating Gammatone Wavelet Cepstral Coefficients (GWCC), Mel Frequency Cepstral Coefficients (MFCC), Pitch, Vocal Tract Frequency, And Spectral Energy; And An Artificial Neural Network (ANN) Classifier Optimized Using A Bio-inspired Hybrid Framework Of The Bat Algorithm And Particle Swarm Optimization (BAT+PSO) [4], [3]. Extensively Evaluated On The Benchmark SUSAS Dataset And A Custom Marathi Speech Database, The Proposed Framework Completely Bypasses Localized Gradient Trapping To Deliver An Outstanding Overall Stress Classification Accuracy Of 84.2% With A Minimal Mean Square Error (MSE) Of 0.0170 [5].

Keywords