Polyphase Quadrature Filters

Background

While working for RCA I started looking at the problem of filtering a digital audio signal into multiple frequency bands, in such a way that

the bands can be recombined to nearly exactly reproduce the original signal.

The total number of samples stays the same. For example, filtering an 800 samples/second stream into 8, 1000 samples/second stream. Previously, quadrature mirror filters (QMF) had been used to split a band into two equal bands with half the sampling rate. Trees of QMF's could therefore be used to split a signal into more than 2 bands.

Also, techniques had been developed to do multiband filtering using frequency transforms (FFT or DCT). However, these techniques required an increase in the total number of samples.

The key contribution of this paper is the development of a computationally efficient way to perform multiband filtering using a single prototype filter and a Discrete Cosine Transform (DCT). By properly phasing the different bands, it is possible to cancel the resulting aliasing, and recover the original signal.

The original application was subband coding of speech signals. A similarity has been noted between this technique and the filter bank used in the MPEG-2 Audio Codec III (the so-called mp3 files). For example, in "A Tutorial on MPEG/Audio Compression" by Davis Pan, he states:

The ISO MPEG audio standard describes a procedure for computing the analysis polyphase filter outputs that is very similar to a method described by Rothweiler

A list of citations of this paper is available at http://citeseer.nj.nec.com/context/609944/0

For my paper I referred to this type of filter bank as a Polyphase Quadrature filter bank, combining the names of the two-band and multiband filter structures. Today, it is more commonly referred to as a pseudo-QMF filter bank. This distinguishes it from the perfect-reconstruction (PR) filterbanks that were later developed by other researchers. As the name implies, PR filterbanks can combine the subbands to exactly recover the input signal, while the pseudo-QMF approach only approximately recovers it.

However, with reasonable filters, the pseudo-QMF approach can achieve reconstruction with 60 dB or more of attenuation of the error components, so they are quite adequate for many practical purposes.

Paper

Copyright information.
The publisher, IEEE, requires the following copyright notice to be posted. For more information, check the IEEE policy (6.2.10).

"© 1983 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE."

To view to download the paper: icassp83_27-2.pdf (327337 bytes)
This PDF file is a scanned image of the published paper.
Optional: click here to send me an email to let me know you downloaded the paper.