Fast Octave Filter Banks

Let's now apply the above technique to the design of an octave filter bank.^11.19 At first sight, this appears to be a natural fit, because it is immediately easy to partition a power of 2 (typical for the FFT size N) into octaves, each having a width in bins that is a power of 2. For example, when N == 8, we have the following stack of frequency-response vectors for the rectangular-window, no-zero-padding, complex-signal case:

H = [ ...
   0   0   0   0   1   1   1   1 ; ...
   0   0   1   1   0   0   0   0 ; ...
   0   1   0   0   0   0   0   0 ; ...
   1   0   0   0   0   0   0   0 ];

Figure 10.31 depicts the resulting filter bank schematically in the frequency domain.^11.20 Thus, H(1,:) is the frequency-response for the top (first) octave, H(2,:) is the frequency-response for the next-to-top (second) octave, H(3,:) is the next octave down, and H(4,:) is the ``remainder'' of the spectrum. (In every octave filter bank, there is a final low-frequency band.

Figure 10.31: Simple octave filter bank for complex signals.

A schematic of the filter-bank implementation using FFTs is shown in Fig.10.30. The division by N called for by ifft(N) is often spread out among the ifft butterfly stages as divisions by 2 (one-bit right-shifts in fixed-point arithmetic) after each of them. For conservation of dynamic range, half of the right-shifts can be spread out among every other forward-FFT butterfly stage [28] as well.

The simple spectral partitioning of Fig.10.31 is appropriate for complex signals--those for which the spectrum is regarded as spanning 0 Hz to the sampling rate. For real signals, we need a spectral partition more like that in Fig.10.32.

Figure 10.32: Simple octave filter bank for real signals.

Unfortunately, the number of spectral samples in Fig.10.32 is 14--not a power of 2. (The previous length 8 complex case maps to length 14 real case because the dc and Nyquist-limit samples do not have complex-conjugate counterparts.) Discarding the sample at the Nyquist limit--number 8 in Fig.10.32--does not help, since that gives 13 samples--still not a power of 2. In summary, there is no obvious way to octave-partition the spectral samples of a real signal while maintaining the power-of-2 condition for each band and symmetrically partitioning positive and negative frequencies.

A real signal can of course be converted to its corresponding ``analytic signal'' by filtering out the negative-frequency components. This is normally done by designing a Hilbert transform filter [133,224]. However, such filters are large-order FIR filters, exactly like we are trying to design! If we design our filter bank properly, we can use it to zero the negative-frequency components.