A method for estimating noise quantity by neural network in speech enhancement system

Abstract

In order to reduce the noise in the noisy speech, it is desirable to change the method of processing according to the intensity of the noise amount in order to reproduce a speech of good quality. Realizing this idea, we applied a three layered neural network which is able to learn the three graded speeches containing three graded white noise. Experimental results demonstrate that the noise amount could be estimated by the neural network, whose inputs are cepstral coefficients. We believe that the improvement of speech signals degraded by noise is fully accomplished by this system.