AG Kommunikationstheorie

Thema:

Design of Filter-Banks with Non-Uniform Frequency Resolution

Abstract:

Many Algorithms for (adaptive) speech and audio processing operate in the short-time frequency domain using an analysis-synthesis filter-bank system. For such systems the use of filter-banks with non-uniform subband-widths which mimic the non-uniform frequency resolution of the human auditory system is beneficial. One flexible option to construct non-uniform filter-banks is by means of an allpass transformation where the delay elements of a uniform nonrecursive filter-bank are replaced by allpass filters.

Recently, a novel numerical design method for such allpass transformed analysis-synthesis-filter-banks has been developed which is based on solving convex optimization problems. The method has been developed for filter-banks based on a discrete fourier transform (DFT) and leads to systems with significantly reduced reconstruction errors compared to related filter-bank designs.

The presentation treats the application of this promising approach to a much wider class of filter-banks. One main part is the application of the considered design method to allpass transformed filter-banks based on a discrete cosine transform (DCT-filter-banks), which have wide application in speech processing. Another part is the design of non-uniform two-channel filter-banks with critical subsampling, which are attractive for coding applications. In both cases a key challenge is to reformulate the designs in such a manner that the obtained optimization problems stay convex since convex optimization problems provide a global optimum and can be solved efficiently.



zurück zur Terminübersicht