Voice profile identification using FFT, multilayer PNN and FFNN approach

The article proposes an approach for recognition of voice profiles for identification of individuals based on obtained FFT spectral characteristics and multilayer artificial neural networks. The approach integrates Probabilistic Neural Networks (PNN) and Feed-Forward Neural Networks (FFNN) based on Deep Learning. Information retrieval procedures were carried out when processing target voice profiles by applying different types of Window Function to the algorithm of Fast Fourier Transform. Four-layer PNN architectures were studied at different levels of the Spread Indicator in Radial-Basis network layer. A comparative analysis was conducted between Deep FFNNs when implementing hyperbolic tangent and log-sigmoid transfer functions in output layers. Quality of classification indicators, respectively Accuracy, Mean-Squared Error (MSE) and Mean Absolute Error (MAE) were assessed. According to the procedures applied, a correct identification of personalized voice profiles was established.