filterbank

A filterbank is a collection of filters designed to partition an input signal into multiple components, each corresponding to a specific frequency band. Filters in a bank pass a chosen range of frequencies while attenuating others, and the outputs—the subbands—can be processed individually or transmitted at lower data rates. Filterbanks appear in both analog and digital form, with digital implementations common in modern systems.

Filterbanks are typically described as analysis filters that decompose a signal, and, when needed, synthesis filters that recombine subbands to reconstruct the original signal. They may be uniform, offering equal bandwidths, or non-uniform, providing varying bandwidths to match perceptual or system requirements. Common architectures include polyphase filter banks, FFT-based banks, and quadrature mirror filter banks, each balancing spectral selectivity and computational efficiency.

Key concepts include decimation and upsampling, which reduce or restore data rate in subbands, and the handling of aliasing and reconstruction errors. Ideal filter banks achieve perfect reconstruction under certain conditions, though practical designs trade perfect reconstruction for finite impulse response length, computational cost, and tolerance to perturbations. Prototype filters or windowing functions are often used to shape the subband filters.

Applications span diverse fields: audio processing (graphic equalizers, subband coding, and perceptual filter banks such as mel-filter banks in speech processing), communications (channelizers in receivers and software-defined radios), and hearing devices. In modern codecs, filter banks are integral components of broader transform-based coding schemes, enabling efficient representation and processing of multi-band signals.