A Fair Play Comparison of Neural Networks Algorithms in Speech Signal Processing

The separation of noise from speech signals has always been a necessary requirement and is being demanded in speech signal processing applications as an important factor to achieve clear speech while communicating. In the last decade, some research on the separation of speech and noise has been published. In previous research work some traditional algorithms were used, such as least mean squares, the nearest neighbour and the quadratic Gaussian algorithms. This paper provides an extensive experimental simulation of speech with and without noise to solve the filtering problem. An FIR digital filter is designed and proposed to train a neural network. The experimental results show that using neural networks in noise separation produce a more robust and powerful separation of speech and noise than other traditional algorithms. Furthermore, the FIR digital filter provides a fast convergence and gives results near the global optimal. The neural networks such as Elman, radial base function and perceptron networks are trained with different training algorithms and compared with the performance of FIR digital filter including its computational complexity. It is found that an algorithm chosen to train the neural network is very important to the final results