data_structure

利用HMM的方法的三种语音识别算法

The routines use a common data structure 'hmm' with fields:


hmm.train.cyc		max number of cycles through data
	  .tol		termination tolerance of likelihood
	  .rdisplay	continous output of free energy values
	  .FrEn         free energy terms for each iteration
	  .Xi           joint probability of past and future states 
	  	  	conditioned on data 
	  .Gamma        probability of states conditioned on data 
	  .Gammasum     expectation of Gamma over time

hmm.data.Xtrain		cell array containing training data
         .T		length of training sequence
		

   .obsmodel	name of observation model
		'Gauss'		- Gaussian
		'LIKE'          - observations are likelihoods
		'Poisson'	- Poisson
		'Dirichlet'	- Discrete 
   .K		dimension of state-space
   .Pi 		expected initial state probability 
   .P		expected state transition probabilities 
   .gmmll	log-likelihood of gmm model used for initialisation
   .mix		Gaussian mixture model trained on same data

   .train.init   	initialisation flag  (1 or 0)
         .obsupdate	update observation model (1 or 0)
         .pupdate	update transition matrix (1 or 0)
	   
   .Dir2d_alpha   posterior 2-D Dirichlet for Tx Probabilities
   .Dir_alpha     posterior Dirichlet for intial state Probabilities 

   .prior.Dir2d_alpha  2-D Dirichlet prior for Tx Probabilities
         .Dir_alpha    Dirichlet prior for initial state Probabilities



For 'Gauss' observation models we also have:
hmm.state(:)
	   .Mu		Expectation of Posterior for mean
           .Cov         Expectation of Posterior for Covariance
	   .Norm_Mu     Posterior for mean: mean (1,dimension(data))
           .Norm_Cov    Posterior for mean: covariance
           .Norm_Prec   Posterior for mean: precision
           .Wish_alpha  Posterior for Covariance: scale parameter
           .Wish_B      Posterior for Covariance: shape matrix
           .Wish_iB     inverse of .Wish_B 
           .prior       priors for each state
                 .Norm_Mu     Prior for mean: mean (1,dimension(data))
                 .Norm_Cov    Prior for mean: covariance
                 .Norm_Prec   Prior for mean: precision
                 .Wish_alpha  Prior for Covariance: scale parameter
                 .Wish_B      Prior for Covariance: shape matrix
                 .Wish_iB     inverse of .Wish_B 
		 .Wish_k      dimension of B

For 'Poisson' observation models we also have:
hmm.state(k)
           .Gamma_alpha	  Posterior for rate: scale parameter
           .Gamma_beta    Posterior for rate: shape parameter
           .prior         priors for each state
                .Gamma_alpha   prior for rate: scale parameter
                .Gamma_beta    prior for rate: shape parameter


For 'Dirichlet' observation models we also have:
hmm.state(k)
           .Dir_alpha     posterior for cell Probabilities 
           .prior       priors for each state
                 .Dir_alpha     prior for cell Probabilities 



For 'LIKE' observation models, there are no extra parameters.