MMSE filter banks with reduced complexity

This paper focuses on subband coding schemes based on critically decimated paraunitary filter banks. An additional network with matrix A is introduced on the decoder side. We show here how to optimize its coefficients jointly with the quantization steps in order to reduce the reconstruction mean square error (MSE) on the output signal due to quantization and filtering. This optimization is performed under bit-rate constraint. Of course, the resulting overall MMSE filter bank (including A ≠ Im) does not allow perfect reconstruction (PR), but the signal-to-noise ratio (SNR) is remarkably better than for the PR solutions. The main advantage of such an approach is to preserve the original structure of the filter bank while improving the SNR. Here, we apply the method to modulated filters which can be implemented at very low cost. Solutions are given by comparing rate-distortion curves for various forms of A, several paraunitary modulated filter banks, and two bit-rate measures.