Speech enhancement using adaptive neuro-fuzzy filtering

This paper presents an adaptive neuro-fuzzy filtering scheme using the Artificial neuro-fuzzy inference system (ANFIS) for noise reduction in speech. The measurable output noisy speech with 5dB SNR level is taken as the contaminated version of the interference to compare with the output data of the filter. The white noise source is taken as the input. With separate sets of input and output vectors formed after subtractive cluster estimation, an initial first-order (Takagi-Sugeno-Kang) TSK Fuzzy Inference System (FIS) is generated. The number of rules and antecedent membership functions of the FIS is determined based on the estimated cluster centres and then uses linear least squares estimation to determine each rule's consequent equations. This function returns the initial FIS structure that contains a set of fuzzy rules to cover the feature space. Finally, the ANFIS hybrid-learning algorithm that combines the recursive least-squares estimation (RLSE) method and the back propagation gradient descent (BP/GD) is applied to determine the premise and the consequent parameters. After training, the ANFIS output (i.e. estimated interference) was determined. Then the estimated information signal is calculated as the difference between the measured signal and the estimated interference. It was noted that without extensive training, the ANFIS could do a fairly good job in adaptive denoising of a speech system with nonlinear characteristics.