% 2008-04-03-ftrclass.diary % Actually try classifying on some real speech data % dpwe@ee.columbia.edu http://www.ee.columbia.edu/~dpwe/e6820/ % Read in the phoneset definition K = labreadkey('timit61.phset'); % Define 6 broad classes covering the 61 phones in the set brdclass=[ones(1,13)*1,ones(1,10)*2,ones(1,7)*3,ones(1,5)*4,[2 2],ones(1,20)*5,[6 6 6 6]]; % 1=stops 2=fricatives 3=nasals 4=liquids 5=vowels 6=silence/gaps % Look at the labeling for one TIMIT file: % load waveform [d,sr] = wavread('mdpk0/sa1.wav'); % load hand-label definitions (t defines times and l defines labels, % indexed into K) [t,l] = labreadlab('mdpk0/sa1.phn',K); subplot(211) specgram(d,128,sr); % Replace the time axis with the phone label boundaries labplotlabs(t,l,K) % Calculate cepstral coefficients for this example using Malcolm Slaney's % routine cp = mfcc(d,16000,100); % Set up a vector of the times corresponding to each column of cp % (100 Hz frame rate, but 256 pt window, so first window is at 0.008 s) tt=0.008:.01:(length(d)/sr-0.018); % Now figure out the labels that go with each feature vector % by 'sampling' the label ranges read from file ll=labsamplabs(tt,t,l); % Using the label values in ll to index into brdclass gives us the % broad class indices (1-6) above subplot(212) plot(brdclass(ll)) axis([0 length(ll) 0 7]) % Lines up with segments in spectrogram. % Use utility functions to build up training and test sets of % features and labels, as above % Define a train data set tdat=[]; tlab=[]; [tdat,tlab]=appenddata('mdpk0/sx153',tdat,tlab,K); [tdat,tlab]=appenddata('mdpk0/sx243',tdat,tlab,K); [tdat,tlab]=appenddata('mdpk0/sx333',tdat,tlab,K); [tdat,tlab]=appenddata('mdpk0/sx423',tdat,tlab,K); [tdat,tlab]=appenddata('mdpk0/sx63',tdat,tlab,K); % And a test set edat=[]; elab=[]; [edat,elab]=appenddata('mdpk0/sa1',edat,elab,K); [edat,elab]=appenddata('mdpk0/sa2',edat,elab,K); [edat,elab]=appenddata('mdpk0/si1053',edat,elab,K); [edat,elab]=appenddata('mdpk0/si1683',edat,elab,K); size(edat) ans = 1131 13 size(tdat) ans = 1453 13 % Try a classification based on the first 8 cepstra [H,O,R]=doclassif(tdat(:,1:8),brdclass(tlab),edat(:,1:8),brdclass(elab)); Ep 1 lr=0.4 frame accuracy on trn = 78.6648% test = 72.5022% Ep 2 lr=0.2 frame accuracy on trn = 77.7013% test = 66.7551% Ep 3 lr=0.1 frame accuracy on trn = 79.3531% test = 71.176% % Do it again [H,O,R]=doclassif(tdat(:,1:8),brdclass(tlab),edat(:,1:8),brdclass(elab)); Ep 1 lr=0.4 frame accuracy on trn = 80.3166% test = 71.9717% Ep 2 lr=0.2 frame accuracy on trn = 80.4542% test = 72.1485% Ep 3 lr=0.1 frame accuracy on trn = 81.1425% test = 70.4686% % Several percent variation in successive runs is not unusual % Check confusion matrix confus(brdclass(elab),R) ans = 58 5 7 0 13 31 17 112 1 1 24 2 5 1 46 0 8 0 15 1 9 17 49 0 8 3 6 21 367 2 52 43 0 1 9 197 % Vowels and fricatives are distinct; silence and stops confused