This paper deals with the application of Higher Order Statistics (HOS) in a recognition system where the speech input is corrupted by noise. Cumulants, Biespectrum and Trispectrum are being used in noisy speech signal processing because of their immunity to noise. In this paper we introduce the use of integrated polispectrum and triespectrum in a recognition system based on a filter bank analysis. The results improves other HOS based methods without a substantial increase of the computational load.