MATLAB Murphy's HMM Toolbox

Question

I am trying to learn HMM GMM implementation and created a simple model to detect some certain sounds (animal calls etc.)

I am trying to train a HMM (Hidden Markov Model) network with GMM (Gaussian Mixtures) in MATLAB.

I have a few questions, I could not be able to find any info about.

1) Should mhmm_em() function be called in a loop for each HMM-state or it is automatically done?

Such as:

 for each state
        Initialize GMM’s and get parameters (use mixgauss_init.m)
    end
    Train HMM with EM (use mhmm_em.m)

2)

[LL, prior1, transmat1, mu1, Sigma1, mixmat1] = ...
                            mhmm_em(MFCCs, prior0, transmat0, mu0, Sigma0, mixmat0, 'max_iter', M);

The last Parameter, should it be the number of Gaussians or a number_of_states-1?

3) If we are looking for Maximum likelihood, then where the Viterbi comes into play?

Say if I want to detect a certain type of animal/human call after training my model with the accoustic feature-vectors that I have extracted, should I still need a Viterbi algorithm in test mode?

It is a little bit confusing me and I would highly appreciate an explanation for this part.

Any comments for the code in terms of HMM GMM logic would also be appreciated.

Thanks

Here is my MATLAB routine;

O = 21;            % Number of coefficients in a vector(coefficient)
M = 10;            % Number of Gaussian mixtures
Q = 3;             % Number of states (left to right)
%  MFCC Parameters
Tw = 128;           % analysis frame duration (ms)
Ts = 64;           % analysis frame shift (ms)
alpha = 0.95;      % preemphasis coefficient
R = [ 1 1000 ];    % frequency range to consider
f_bank = 20;       % number of filterbank channels 
C = 21;            % number of cepstral coefficients
L = 22;            % cepstral sine lifter parameter(?)

%Training
[speech, fs, nbits ] = wavread('Train.wav');
[MFCCs, FBEs, frames ] = mfcc( speech, fs, Tw, Ts, alpha, hamming, R, f_bank, C, L );
cov_type = 'full'; %the covariance type that is chosen as ҦullҠfor gaussians.
prior0 = normalise(rand(Q,1));
transmat0 = mk_stochastic(rand(Q,Q));
[mu0, Sigma0] = mixgauss_init(Q*M, dat, cov_type, 'kmeans');

mu0 = reshape(mu0, [O Q M]);
Sigma0 = reshape(Sigma0, [O O Q M]);
mixmat0 = mk_stochastic(rand(Q,M));
[LL, prior1, transmat1, mu1, Sigma1, mixmat1] = ...
mhmm_em(MFCCs, prior0, transmat0, mu0, Sigma0, mixmat0, 'max_iter', M);

%Testing
for i = 1:length(filelist)
  fprintf('Processing %s\n', filelist(i).name);
  [speech_tst, fs, nbits ] = wavread(filelist(i).name);
  [MFCCs, FBEs, frames ] = ...
   mfcc( speech_tst, fs, Tw, Ts, alpha, hamming, R, f_bank, C, L);
  loglik(i) = mhmm_logprob( MFCCs,prior1, transmat1, mu1, Sigma1, mixmat1);
end;
[Winner, Winner_idx] = max(loglik);

Answer 1

1) No, EM estimates the model as a whole after you initialized it with kmeans. It doesn't estimate states separately.

2) Neither, last parameter in your code is the value of 'max_iter', it is the number of iterations of EM. Usually it's something around 6. It should not be M.

3) Yes, you need Viterbi in test mode.