Stress Recognition Using Sound Analysis, k-NN, Decision Tree and Artificial Intelligence Approach

Abstract

The paper presents an approach for stress recognition based on registered sound parameters LZE, LZeq, LZF and LZS in different speech levels in working environment. The approach combines k – nearest neighbors (k-NN) method in Euclidean, Cityblock, Minkowski and Chebychev metric distances, Decision tree (DT) method with CART algorithm and artificial neural networks (ANN) with Levenberg-Marquardt (LM) and Scaled Conjugate Gradient (SCG) algorithms. According to k-NN method a maximum accuracy of 93.99 % with minimum parameter k = 3 for Cityblock distance have been registered. There has been established a fourth optimum level of nodes pruning from the structure for multiple choice of the classification group by used on DT method with achieved accuracy of 99.05 %. In investigation of LM algorithm during training of networks with purelin, tansig and logsig transfer functions has been achieved identical accuracy of 99.99 %. ANN architecture with tansig output activation function has been selected With regard to a minimal indication of Mean Squared Error (MSE) indicator 0.0064 in 11 hidden neurons. By using of artificial intelligence (AI) during SCG training was synthesized a model for correct speech recognition with level accuracy 100.00 %.