% 2009-02-19-lpc.diary
% Examples of using LPC functions
% Load speech soundfile
[d,sr]=wavread('mpgr1_sx419.wav');
% Downsample to 8 kHz (LPC works better with a limited bandwidth)
d = resample(d,1,2);
sr = sr/2;
soundsc(d,sr);
% 8th-order LPC modeling
[a,g,e] = lpcfit(d,8);
% Noise-excited resynthesis
rd = lpcsynth(a,g);
% Listen
soundsc(rd,sr)
% Compare spectrograms
subplot(211)
specgram(d,256,sr);
subplot(212)
specgram(rd,256,sr);
% Clearly the same broad shape, but without pitch structure.
% Residual-excited resynthesis (i.e. using the 'e' argument)
rde = lpcsynth(a,g,e);
soundsc(rde,sr)
% Perceptually identical to original
% What is the excitation signal like?
% It should contain everything *not* in the noise-excited reconstruction
soundsc(e,sr)
% Weird, flat power in time and frequency, still intelligible!
specgram(e,256,sr);
% Interesting that formants still visible, resulting from slight
% modeling mismatch in the LPC
% Compare original waveform with residual
plot(d)
% Choose a portion straddling fricative-vowel in "football"
subplot(311)
plot(d(2000+[1:1000]))
subplot(312)
plot(rde(2000+[1:1000]))
% Residual-excited resynth is close but not identical due to
% the overlap-adding of adjacent frames
subplot(313)
plot(e(2000+[1:1000]))
% Residual is basically noise during fricatives or silence;
% clearly shows pitch pulses during voicing
% Let's just check the model during this vowel
% Frames occur every 128 samples, so middle of vowel (sample 2500)
% is around frame 20... Look at the gain (energy) terms to
% find peak
g(17:23)
%ans =
% 0.0015
% 0.0009
% 0.0053
% 0.0071
% 0.0037
% 0.0003
% 0.0000
% g(20) is the peak. Let's see the spectrum: reconstruct the polynomial
a20 = a(20,:);
g20 = g(20);
% see the filter response (scale up by energy of one window's excitation)
[h,w]=freqz(g20*sqrt(256), a20);
subplot(211)
plot(w/pi, 20*log10(abs(h)));
% Compare to poles of the IIR filter
subplot(212)
zplane(roots(a20))
% poles close to unit circle corresponding to spectral peaks.
% Compare to spectral slice at that point
% (Include pre-emphasis in spectrogram, to match actual modeled signal)
sg=specgram(filter([1 -0.9], 1, d),256,sr);
subplot(212)
plot([0:128]/128, 20*log10(abs(sg(:,20))))
axis([0 1 -40 0])
grid
% LPC traces/approximates general trend of actual spectrum: superimpose it
hold on
plot(w/pi, 20*log10(abs(h)), '-r')
hold off