This paper presents a study on the implementation of a Normalized Least Mean Square (NLMS) algorithm on a TI TMS320C54x DSP. The architecture of the DSP and the instruction set are designed to optimize the computation of usual signal processing algorithm. This fixed point DSP family is especially suited for wireless applications and allows efficient implementations of advanced speech processing applications such as Echo Cancellation. But the specific mapping of the algorithm on the DSP changes the convergence properties of the classical NLMS based echo canceller. Therefore NLMS equations need to be adapted to DSP implementation. This paper shows that this DSP implementation reduces the computational load of the classical NLMS by 66% while maintaining good convergence stability and high performances close to usual floating point implementation.
展开▼