Comparison of classification techniques for speechaudio applications

Abstract

This classification of speech vs. noise or speech vs. music can be used in speech/audio signal processing applications as an important part to achieve a lower bit-rate or to enhance the performance of coding in multimedia communications. In recent years, some research on the classification has been published. In the previous work some traditional classification algorithms were used, such as linear Least Mean Squares, the Nearest Neighbour and the Quadratic Gaussian algorithms. Some of the previous work also used machine learning software and neural networks with basic training algorithms. This thesis provides an extensive experimental simulation of the speech classification problem. In this thesis, the Extended Kalman Filter algorithm is proposed to train a neural network classifier. Our experiments show that using neural networks in speech/noise/music classification produces a more robust and powerful classification than other traditional algorithms. Furthermore, the Extended Kalman Filter provides a fast convergence and gives results near the global optimal. Our results show that the learning algorithm chosen to train the neural network is very important to the final results. Therefore, the neural network classifier trained with the Extended Kalman Filter is compared with the traditional classification methods but also with other neural classifiers trained with different learning algorithms. The experiments performed for the thesis are clean speech vs. noise classification, clean speech vs. music classification and active/inactive speech classification of noisy speech.