Speech enhancement is one of the important preprocessing techniques required for any speech processing applications. This work represents dual channel speech enhancement where the desired signal is available and the input noisy speech is enhanced based on the reference signal, this method of speech enhancement can be used in robots where the machine can recognize the comments given by human. In this paper noisy and the desired speech signals are dual transformed using discrete cosine transform and Hadamard transform and applied to the adaptive filter using Normalized Least Mean Square algorithm. In NLMS the step size parameter is varying based on the input signal rather in LMS the step size is constant. This variable step size will lead to fast convergence of noisy speech towards desired speech and the enhanced signal gives better performance compared to conventional LMS algorithm. The performance analysis is done through various subjective and objective measures.