Low complexity noise power estimator for speech enhancement implemented on a dsPIC

In speech processing, the Signal-to-Noise Ratio (SNR) of the signal is an important feature. There are methods to reduce the noise contained into the speech which allow to obtain better results of the processing carried out. In this work a set of adaptive filtering methods are studied, with a deep analysis of the noise power estimators used to carry out the speech enhancement. Two baseline estimators are studied and a third estimator, which has lower computational complexity than the others, is presented. Finally, a set of implementations are performed in both MATLAB and a low cost hearing aid device based on the dsPIC33EP256MU806 from Microchip. A set of objective experiments and experimental measures are developed to verify the performance of the system.