fast_mvg.c File Reference

Gaussian with diagonal covariance - fast evaluation. More...

Data Structures
struct	SprHmmFGset

Enumerations
enum	{ SPR_HMM_FGMO_S_AV, SPR_HMM_FGMO_S_MAX, SPR_HMM_FGMO_G_MAX }

Functions
void	spr_fmvg_dump (SprStream *fd, const SprHmmFGset *gset)
int	spr_check_keys (const char **key_desc, SprKeySet *keys, SprMsgId *routine, const char *fname)
float *	spr_fmvg_mg_prob (const float *x, SprAcmod *restrict gset, float *restrict pvec)
float *	spr_fmvg_output_prob (const float *x, SprAcmod *restrict gset, float *restrict oprob)
float *	spr_fmvg_output_prob_mo (const float *x, SprHmmFGset *restrict gset, float *restrict oprob)
float *	spr_fmvg_log_oprob (SprAcmod *restrict gset)
SprAcmod *	spr_fmvg_read (const char *hmm_fname, const char *mvg_fname, const char *sel_fname)

Variables
const char *const	spr_hmm_str_fgmo []
const SprHmmFGset	spr_hmm_empty_fgset
const char *	spr_check_hmm_keys []
	required keys in an HMM-file and their correct values More...
const char *	spr_check_mvg_keys []
	required keys in an MVG-file and their correct values More...
const char *	spr_check_sel_keys []
	required keys in an SEL-file and their correct values More...
const SprCmdOptDesc	spr_cwr_od_am_std []
const SprCwrSRMDesc	spr_cwr_srm_desc_am_std

Detailed Description

Gaussian with diagonal covariance - fast evaluation.

A more efficient (but less straight-forward) implementation of gaussian mixtures. The data organisation is also optimized for bottom-up gaussian selection (FRoG). The new implementation also allow some simple schemes to combine multiple observations.

Options for loading a setup with multiple observations:

[AM_options]
Noversample=<I32:val>(1)
number of observations per frame
combine=<SAv/SMax/GMax>
combine score by taking the average state score (SAv), the best state score (SMax), or the best score per Gaussian before calculating the state scores (GMax)

Author

Kris Demuynck

Date

24 Feb 1999

Revision History:

24/02/99 - KD

creation

03/12/03 - KD

added multiple observations per vector