Female Voice Recognition Using Artificial Neural Networks and MATLAB Voicebox Toolbox

Abstract

Voice and speaker recognition performances are measured based on the accuracy, speed and robustness. These three key performance indicators are primarily dependent on voice feature extraction method and voice recognition algorithm used. This paper aims to discuss various researches in speech recognition that has yielded high accuracy rates of 95% and above. The extracted MFCCs from MATLAB Voicebox toolbox were used as inputs to the multilayer Artificial Neural Networks (ANN) for female voice recognition algorithm. This study explored the recognition performance of the neural networks using variable number of hidden neurons and layers, and determine the architecture that would provide the optimum performance in terms of high recognition rate. MATLAB simulation resulted to a training and testing recognition rate of 100.00% when using 3-hidden-layer neural network from speech samples of a single-speaker, and highest training recognition rate of 98.11% and testing recognition rate of 87.20% when using 4-hidden-layer neural network from speech samples of several speakers. When tested with homonyms, the best recognition rate was 75.00% from a 3-hidden-layer neural network trained from a single-speaker, and 81.91% from a 4- hidden-layer neural network trained from multiple speakers. The deviation in recognition rates were primarily attributed to the variations made in the number of input neurons, hidden layers, and neurons of the speech recognition neural network.