gmm_util.H File Reference

Miscellaneous routines for manipulating GmmSet's. More...

#include "util.H"

Go to the source code of this file.

Classes
class	GmmStats
	GMM statistics for reestimating a GmmSet. More...
Functions
void	init_simple_gmms (GmmSet &gmmSet, unsigned gmmCnt, unsigned dimCnt)
	Initializes gmmSet to contain gmmCnt GMM's of dimension dimCnt.
void	split_gmms (const GmmSet &srcGmmSet, GmmSet &dstGmmSet, const map< string, string > &params=ParamsType())
	Splits each Gaussian in each GMM in srcGmmSet in two along the direction of the variance, and places result in dstGmmSet.
void	expand_gmms_ci_to_cd (const GmmSet &srcGmmSet, GmmSet &dstGmmSet, const vector< int > &phoneGmmCounts)
	Duplicates each (context-independent) GMM in srcGmmSet for each of its (context-dependent) variants and places result in dstGmmSet.

---

Detailed Description

Miscellaneous routines for manipulating GmmSet's.

---

Function Documentation

void init_simple_gmms	(	GmmSet &	gmmSet,
		unsigned	gmmCnt,
		unsigned	dimCnt
	)

Initializes gmmSet to contain gmmCnt GMM's of dimension dimCnt.

Each GMM is taken to have a single component Gaussian with mean 0 and variance 1 for each dimension.

void split_gmms	(	const GmmSet &	srcGmmSet,
		GmmSet &	dstGmmSet,
		const map< string, string > &	params = ParamsType()
	)

Splits each Gaussian in each GMM in srcGmmSet in two along the direction of the variance, and places result in dstGmmSet.

Divides mixture weight of each Gaussian component in two as well. Number of standard deviations to perturb the mean in each dimension is determined by perturb parameter in params. The min_perturb parameter in params controls the minimum distance to perturb each mean. Handles case where Gaussians are shared between GMM's correctly.

void expand_gmms_ci_to_cd	(	const GmmSet &	srcGmmSet,
		GmmSet &	dstGmmSet,
		const vector< int > &	phoneGmmCounts
	)

Duplicates each (context-independent) GMM in srcGmmSet for each of its (context-dependent) variants and places result in dstGmmSet.

The vector phoneGmmCounts holds the number of times to duplicate each source GMM. If Gaussians are shared between GMM's in input, will no longer be shared in output.