Probabalistic Neural Network on Higher Order Spectrum Classification

Many tasks in classification and recognition system need to be performed in real time. While there are difficulties due to the complexity of considered pattern, i.e., as the pattern becomes more complex the training a set of patterns is insufficient to represent what will arise during tests. In our previous study, we characterized the Higher order spectrum (HOS) pattern property and its classification strategy. The HOS pattern is prefered to be used for some classification tasks, as for their interesting characteristic, i.e, higher dimensionality, Gaussian noise removal, and phase relation quantification. An appropriate feature extraction mechanism is certainly needed in order to quantize this HOS input pattern before classification. However, for a real-time classification task, this well-developed feature extraction process should be adhered by a robust and fast classifier. In this study, we present the use of Probabilistic Neural Network(PNN) as Neural Classifier and compared it with a conventional Back Propagation(BP) learning scheme. Both memorization and generalization comparison of those networks are discussed as for developing a robust and fast speaker identification in a noisy environment. It was shown that even using only a small size codebook, PNN can classify bispectrum pattern above 88% success rate and 36 times faster than BP.