segReest.c File Reference

Reestimation for uniform trajectory segmental HMMs. More...

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <float.h>
#include "segCommon.h"
#include "segStdLib.h"
#include "segReest.h"

Defines
#define	GAMMA_DUR_MODEL
	Defining this means the gamma function is used to approximate the duration distribution.
#define	DUR_PROB_OFFSET   0.0
	This is used to avoid the duration model outputting BIGNUMs.
#define	DUR_SD_FLOOR   (1/sqrt(2))
	Floor of variance for the duration model.
Functions
int	segIncrementLinearHMMAccumulators (SEG_STATE **stateId, int *stateSeq, SEG_SYNTAX *syn, double **spData, int sT, int eT, int maxDur)
	Increment the reestimation bins, according to the current utterance alignment.
int	segInitialiseHMMAccumulators (SEG_MODEL_SET *hmmSet)
	Initialise reestimation accumulators for the entire set of models.
int	segReestSegModels (SEG_MODEL_SET *hmmSet, SEG_CONFIG *config)
	Reestimate linear segmental-HMM parameters.

---

Detailed Description

Reestimation for uniform trajectory segmental HMMs.

Date:

12.05.2006

05.07.2007 (documented)

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

This code was originally created on the basis of Martin Russell & P Jackson's SEGVit Ver.3.4.

Revision

---

Function Documentation

int segIncrementLinearHMMAccumulators	(	SEG_STATE **	stateId,
		int *	stateSeq,
		SEG_SYNTAX *	syn,
		double **	spData,
		int	sT,
		int	eT,
		int	maxDur
	)

Increment the reestimation bins, according to the current utterance alignment.

The function uses the optimal state sequence (i.e., the current utterance alignment) to add each acoustic vector to the accumulator of the corresponding HMM state on the optimal state sequence Linear trajectory model (i.e. slope and mean value).

Parameters:

[in,out]	stateId	Array of state IDs of the best state-sequence
[in]	stateSeq	The best state sequence
[in]	syn	Syntax
[in]	spData	Acoustic feature vectors of speech data
[in]	sT	Start time of the utterance
[in]	eT	End time of the utterance
[in]	maxDur	Maximum segment duration

Return values:

	RTN_ERROR	Memory allocation error
	NO_ERROR	Successfully executed

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

The accumulation is carried out based on the optimal state sequence from training.

int segInitialiseHMMAccumulators

(

SEG_MODEL_SET *

hmmSet

)

Initialise reestimation accumulators for the entire set of models.

The function initialises reestimation accumulators for the entire set of models.

Parameters:

[in,out]

hmmSet

Model set

Returns:

NO_ERROR Successfully executed

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

The following functions are called internally: initTransProbs(), initStates(), initTrjModels() and initDurModels().

See also:

initTransProbs(), initStates(), initTrjModels() and initDurModels().

int segReestSegModels	(	SEG_MODEL_SET *	hmmSet,
		SEG_CONFIG *	config
	)

Reestimate linear segmental-HMM parameters.

The function reestimates the segmental HMMs using the accumulated values in the reestimation bins.

Parameters:

[in,out]	hmmSet	Model set
[in]	config	Configuration parameters

Return values:

	NO_ERROR	Successfully executed
	RTN_ERROR	Too small number of samples for the model

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

The following functions are called internally: modelOccupancy(), reestTransProbs(), reestTrjModels(), durModelOccupancy(), and reestGammaDistDurModels() or reestDiscreteDurModels().

See also:

modelOccupancy(), reestTransProbs(), reestTrjModels(), durModelOccupancy(), reestGammaDistDurModels() and reestDiscreteDurModels()

---