segDecoder.h File Reference

Header file for generic Viterbi-type Segmental-HMM decoder. More...

#include "segSyntax.h"

Go to the source code of this file.

Data Structures
struct	segDpItem
	structure to store decoding results for each node, which are used during trace-back More...
struct	segPruningParameter
	pruning parameters More...
struct	segDecoderParameter
	decoder parameters More...
Defines
#define	VOID   -1
	special value for path pointers of node in trellis
#define	NO_PRUN   0
	no specified pruning method
#define	THRESH_PRUN   1
	beam pruning with probability threshold
#define	N_BEST_PRUN   2
	N best nodes to be maintained.
Typedefs
typedef segDpItem	SEG_DP_ITEM
	structure to store decoding results for each node, which are used during trace-back
typedef segPruningParameter	SEG_PRUN_PARAM
	pruning parameters
typedef segDecoderParameter	SEG_DEC_PARAM
	decoder parameters
Functions
double	segViterbiDecoder (SEG_DP_ITEM ***, SEG_SYNTAX *, double **, int *, int, int, int, SEG_DEC_PARAM *)
	Generic Viterbi-type decorder for segmental HMMs.
int *	segRecoverStateSeq (int *, SEG_DP_ITEM **, SEG_SYNTAX *, int)
	Recover the optimal state sequence from the path record of node.
SEG_ANN_DATA *	segRecoverRecogSeq (SEG_DP_ITEM **, SEG_SYNTAX *, int, int)
	Recover the model labels and their start and end times.
SEG_STATE **	segRecoverStateIds (int *, SEG_SYNTAX *, int)
	Recover the sequence of state identities from a state sequence.
int	segGetDurationLimits (int *, int *, SEG_MODEL_SET *, SEG_SYNTAX *, int)
	Find the minimum and maximum possible durations for a given syntax.
int	segFreeDecoder (SEG_DP_ITEM **, int)
	Free memory space storing all the paths in the trellis.
void	segFreeStateSeq (int *)
	Free memory space of state sequence.

---

Detailed Description

Header file for generic Viterbi-type Segmental-HMM decoder.

Date:

21.06.2007 (revised)

Author:

Yoshinori Shiga, Centre for Vision, Speech and Signal Processing, University of Surrey

Revision

---

Function Documentation

int segFreeDecoder	(	SEG_DP_ITEM **	dpHist,
		int	numStates
	)

Free memory space storing all the paths in the trellis.

Parameters:

[in]	dpHist	Paths in the trellis for back-tracing
[in]	numStates	Number of states

Return values:

NO_ERROR

Successful execution

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

void segFreeStateSeq

(

int *

stateSeq

)

Free memory space of state sequence.

Parameters:

[in]

stateSeq

State (index) sequence

Returns:

No return value

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

int segGetDurationLimits	(	int *	minUttDur,
		int *	maxUttDur,
		SEG_MODEL_SET *	models,
		SEG_SYNTAX *	synt,
		int	eStates
	)

Find the minimum and maximum possible durations for a given syntax.

Parameters:

[in]	minUttDur	Minimum utterance duration
[in]	maxUttDur	Maximum utterance duration
[in]	models	Model set
[in]	synt	Syntax
[in]	eStates	Number of emitting states in the syntax

Return values:

	RTN_ERROR	Inconsistency in the models
	NO_ERROR	Successfully executed

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

This function finds the minimum and maximum possible durations for a given syntax, based on passing through all emitting states in sequence. Thus, the minimum duration is equal to the number of emitting states' maximum duration. At present, this function does not take account of any skips or self loops, but the maximum duration of -1 is reserved for models of possibly infinite duration.

SEG_ANN_DATA* segRecoverRecogSeq	(	SEG_DP_ITEM **	dpMat,
		SEG_SYNTAX *	syn,
		int	numVecs,
		int	numLabs
	)

Recover the model labels and their start and end times.

This function traces back through the DP matrix to obtain the model labels and their start and end times (in frame), ready to write them out in the master label file.

Parameters:

[in]	dpMat	Paths in the trellis for back-tracing
[in]	syn	Syntax
[in]	numVecs	Number of frames in the sequence
[in]	numLabs	Number of models in the sequence

Return values:

	NULL	Memory allocation error
	otherwise	Pointer to the array of struct annData variables each of which contains a label, and start and end times.

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

The first start time is always one but, for classification, the other times are as in the input label file rounded down to the nearest frame.

The purpose of this function is to create a structure in memory which corresponds to the label file which corresponds to the optimal state sequence. In other words, the i-th entry of this array consists of a structure which specifies the start time, end time and label of the i-th label (i.e., model) in the optimal decoding. This structure is the basis of the output for recognition.

SEG_STATE** segRecoverStateIds	(	int *	stateSeq,
		SEG_SYNTAX *	syn,
		int	numVecs
	)

Recover the sequence of state identities from a state sequence.

This function returns the state identities (as pointers to states) of the best state sequence obtained by segRecoverStateSeq().

Parameters:

[in]	stateSeq	State sequence
[in]	syn	Syntax
[in]	numVecs	Number of frames in the utterance

Return values:

	NULL	Memory allocation error
	otherwise	Pointer to the array of struct stateInfo pointers

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

The purpose of this function is to create an array, one entry for each observation vector, such that the t-th entry is a pointer to the HMM state which corresponds to the syntax state at this point on the optimal state sequence. This is useful for re-estimation because it sets up a correspondence between each acoustic vector and the HMM state which it will contribute to for reestimation.

The previous fuction, segRecoverStateSeq(), recovers the indices of the optimal state sequence relative to the syntax. This is stored as an array of ints. The purpose of this function is to convert this sequence of ints into a sequence of pointers to states of actual models.

int* segRecoverStateSeq	(	int *	numLabs,
		SEG_DP_ITEM **	dpHist,
		SEG_SYNTAX *	syn,
		int	numVecs
	)

Recover the optimal state sequence from the path record of node.

This function is used to find the best state sequence by tracing back through the DP matrix with path record of each node in the trellis.

Parameters:

[in,out]	numLabs	Number of models in the sequence recovered
[in]	dpHist	Paths in the trellis for back-tracing
[in]	syn	Syntax
[in]	numVecs	Number of frames in the utterance

Return values:

	NULL	Memory allocation error
	Otherwise	Pointer to the array of state sequence recovered

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

The number of states in the recovered sequence is equal to the number of frames in the utterance. Each element in the sequence is the state number (index) of the corresponding emitting node, and all null states are discarded. This is used for training the model.

See also:

segRecoverStateIds(), segRecoverRecogSeq()

double segViterbiDecoder	(	SEG_DP_ITEM ***	dpHist,
		SEG_SYNTAX *	synt,
		double **	spData,
		int *	breakPt,
		int	sT,
		int	eT,
		int	maxDur,
		SEG_DEC_PARAM *	decParam
	)

Generic Viterbi-type decorder for segmental HMMs.

Parameters:

[out]	dpHist	Paths in the trellis for back-tracing
[in]	synt	Syntax
[in]	spData	Acoustic feature vectors
[in]	breakPt	Break point attributes
[in]	sT	Utterance start time (frame #)
[in]	eT	Utterance end time (frame #)
[in]	maxDur	Maximal segment-duration (Dmax)
[in]	decParam	Decoder parameters

Return values:

	RTN_ERROR	Memory allocation error
	Otherwise	The best cost (minProb) computed

Author:

Yoshinori Shiga @ CVSSP, University of Surrey

Note:

The following functions are called internally: allocDpHistory(), initDecoder(), preCostPartition(), firstPropagator(), selector(), incSnpIdx(), propagator(), prunStartNode(), prunEndNode(), lastPropagator(), lastSelector(), and freeAllNode().

See also:

allocDpHistory(), initDecoder(), preCostPartition(), firstPropagator(), selector(), incSnpIdx(), propagator(), prunStartNode(), prunEndNode(), lastPropagator(), lastSelector() and freeAllNode().

---