SprHmmState Struct Reference

Data Fields
int *	mixdim
float **	weight
SprHmmGndx **	state2mvg
float *	trans
int	sum_ct
int *	trans_ct
	Counts for updating transitions probabilities (for training). More...
float **	weight_ct
	Counts for updating the weights (during training). More...
float *	bias

Detailed Description

Structure describing one single HMM state, modelled with a product (over nmix, the number of separate data streams) of weighted mixtures of gaussians.

Field Documentation

int* SprHmmState::mixdim

The dimension (size) of the mixtures (for each data stream) that model this state.

float** SprHmmState::weight

Weights for the gaussians in the product of mixtures that models this state, e.g. weight[1][5] is the weight for sixth gaussian in the mixture for the second data stream.

SprHmmGndx** SprHmmState::state2mvg

Index of each gaussian the weights refer to. Only for reduced SC-HMMs: for full SC-HMMs this is [0:1:nmix], for CD-HMMs [statenr*nrgauss_per_state:1:nrgauss_per_state]. E.g. state2mvg[1][5] is the index of sixth gaussian in the second set of gaussians (for the second data stream). It corresponds to the weight above.

float* SprHmmState::trans

Transition probabilities. In sequence: probability for self loop, for transition to next state, for skip or transition to second next state, ...

int SprHmmState::sum_ct

The total count (of data points) on which the trans_ct and weight_ct are based.

int* SprHmmState::trans_ct

Counts for updating transitions probabilities (for training).

float** SprHmmState::weight_ct

Counts for updating the weights (during training).

float* SprHmmState::bias

The bias (for each of the data streams) one wants to put on the observation probabilities (always linear, not logarithmic).