Free Books

Spectral Rotation of Real Signals

Note that if we rotate the spectrum of a real signal by half a bin, we obtain $ N/2$ positive-frequency samples and $ N/2$ negative-frequency samples, with no sample at dc or at the Nyquist limit. This is typically desirable for audio signals because dc is inaudible and the Nyquist limit is a degenerate point of the spectrum that, for example, cannot have a phase other than 0 or $ \pi$ . If $ N$ is a power of 2, then so is $ N/2$ , and the octave-band partitioning of the previous subsection can be applied separately to each half of the spectrum:

N = 32;  x = randn(1,N); % Specific example
LN = round(log2(N)); % number of octave bands
shifter = exp(-j*pi*[0:N-1]/N); % half-bin
xs = x .* shifter; % apply spectral shift
X = fft(xs,N); % project xs onto rotated basis
XP = X(1:N/2); % positive-frequency components
XN = X(N:-1:N/2+1); % neg.-frequency components
YP = dcells(XP); % partition to octave bands
YN = dcells(XN); % ditto for neg. frequencies
YPe = dcells2spec(YP); % unpack "dyadic cells"
YNe = dcells2spec(YN); % unpack neg. freqs
YNeflr = fliplr(YNe);  % undo former flip
ys = ifft([YPe,YNeflr],N,2);
y = real(ones(LN,1)*conj(shifter) .* ys);
% = octave filter-bank signals (real)
yr = sum(y); % filter-bank sum (should equal x)
yrerr = x - yr;
  'Total filter-bank sum L2 error = %0.2f %%',...
In the above listing, the function dcells11.21simply forms a cell array in matlab containing the spectral partition (``dyadic cells''). The function dcells2spec11.22is the inverse of dcells, taking a spectral partition and unpacking it to produce a usual spectrum vector.

Next Section:
Improving the Octave Band Filters
Previous Section:
Fast Octave Filter Banks