Abstract: Background noise acoustically adding to speech can degrade the performance of digital environments for applications such as speech comperession and recognition. In this paper we are primarily interested in those situation for which it is known that the noise distribution is nearly normal.We take the noise probability density function (p. d. f) to be the form.As f(x)=(1-ε)g(x)+εh(x) where g(x) is the normal p.d.f,A(x) is an arbitrary symmetric p.d.f,and ε is constant where 0<ε<1. By means of robust estimation,we give the construction of robust M-estimation adding effective function in linear Predictive residue. In ordar to reduce the deviation of estimated parameters of noisy speech, we present a procedure for estimating the robust M and give the experiment data and results.

Key words: Voice recognition, Natural longuage, Speech signal processig.