FIR Digital Filter Banks for Speech Analysis

Filter banks are used to perform short-time spectrum analysis in a variety of speech processing systems. 1-4 Typically, a set of bandpass filters is designed so that a desired portion of the speech band is entirely covered by the combined passbands of the filters composing the filter bank. The outputs of the bandpass filters therefore are considered to be a time-varying spectrum representation of the speech signal. If special care is taken in the design of the bandpass filters, it is possible to reconstruct a very good approximation to the input speech by simply adding together the outputs of the bandpass filters. 5 This is the basic principle of a variety of vocoder systems. Since the bandpass filters are linear systems, we can characterize the behavior of such filter banks by considering the composite frequency response when all the outputs are added together. Since, ideally, the output should be equal to the input, then we desire that the composite frequency response have constant magnitude and linear phase in the desired band of frequencies. This criterion, together with specifications on the desired bandwidths of the individual frequency channels, forms a meaningful basis for the design of filter banks for speech analysis. An earlier paper 5 showed that careful attention to the relative phases between channels is important in achieving a flat composite frequency response. That paper, which was concerned primarily with filter banks composed of infinite impulse response (iir) digital filters, described a method of obtaining flat composite frequency response by a relatively 531