Perfect-reconstruction Biorthogonal Cosine-modulated Filter Banks with Fixed-point Arithmetic

This paper describes the implementation of biorthogonal cosine-modulated filter banks on fixed-point arithmetic digital signal processors. The proposed implementation has the property that the overall filter bank keeps the perfect reconstruction property despite coefficient quantization, overflow, and rounding of intermediate results. The realization of the prototype filter is based on a factorization into zero-delay and maximumdelay matrices. We demonstrate how the frequency selectivity of the filter bank and the coding gain changes with the available wordlength of the fixed-point implementation and the dynamic range of the input signal. For speech signals it turns out that overflow and rounding errors hardly affect the frequency selectivity of the filters if the input signal uses only 75% of the available dynamic range.