Segmenter Class Reference

This class can determine which cluster a certain stream of feature vectors is closest to. The class Train_Segmenter can train these clusters. More...

Inheritance diagram for Segmenter:

[legend]

List of all members.


Public Member Functions
	Segmenter ()
	Segmenter (FILE *inFile)
	~Segmenter ()
float	segmentFeaturePool (int inputSegID, int inputLabelID, int outputSegID)
void	setFeaturePool (FeaturePool *fp)
int	getNumberOfModels () const
Protected Member Functions
void	createLexicalTree (int minNumberOfFrames, int forEachFrames1=NULL, double forEachFrames2=NULL)
float	createSegments (int segID, SegmentationAdmin *admin)
Protected Attributes
int	numberOfClusters
FeaturePool *	featurePool
Vector **	discrTrain
Vector **	discrTrainMask
int	discrTrainLen

Detailed Description

This class can determine which cluster a certain stream of feature vectors is closest to. The class Train_Segmenter can train these clusters.

Constructor & Destructor Documentation

Segmenter::Segmenter ( )

This constructor initialises the cluster parameters and sets the LexicalTree parameters for speaker clustering (defined in standard.h).

References discrTrain, discrTrainMask, numberOfClusters, LexicalTree::overwritePrunePars(), and LexicalTree::overwriteWeightPars().

Here is the call graph for this function:

Segmenter::Segmenter ( FILE * inFile )

This constructor initialises the cluster parameters and sets the LexicalTree parameters for speaker clustering (defined in standard.h). It also loads an existing clustering from inFile.

References discrTrain, discrTrainMask, WriteFileLittleBigEndian::freadEndianSafe(), numberOfClusters, LexicalTree::numberOfPhones, LexicalTree::overwritePrunePars(), LexicalTree::overwriteWeightPars(), and LexicalTree::phoneModels.

Here is the call graph for this function:

Segmenter::~Segmenter ( )

The destructor clears any used memory.

References discrTrain, discrTrainLen, discrTrainMask, LexicalTree::numberOfPhones, and LexicalTree::phoneModels.

Member Function Documentation

void Segmenter::createLexicalTree	(	int	minNumberOfFrames,
		int *	forEachFrames1 = `NULL`,
		double *	forEachFrames2 = `NULL`
	)			`[protected]`

Given the current number of models, create the lexical tree that will be used to align the data.

References LexicalNode::compressedTreeIndex, LexicalNode::contextKey, LexicalNode::contextNext, LexicalNode::contextPrev, LexicalTree::currentlyAligning, LexicalTree::deleteTree(), LexicalNode::depth, LexicalTree::endOfSentenceWord, LexicalTree::grammarStart, LexicalTree::initialiseSystem(), LexicalNode::inputToken, LexicalNode::modelID, LexicalNode::next, LexicalNode::nextTree, LexicalTree::numberOfPhones, LexicalNode::parallel, LexicalTree::phoneModels, PhoneModel::resetPhoneAdmin(), PhoneModel::setSilTrans(), LexicalTree::startOfSentenceWord, LexicalNode::tokenSeq, LexicalNode::tokenSeqLength, and LexicalNode::wordID.

Referenced by Train_Segmenter::createOverlapTree(), Train_Segmenter::getOverlap(), ShoutSegment::ShoutSegment(), and Train_Segmenter::train().

Here is the call graph for this function:

float Segmenter::createSegments	(	int	segID,
		SegmentationAdmin *	admin
	)			`[protected]`

Uses decode result to create some segments, (part of) the pool should have been processed.

References FeaturePool::addSegment(), FeaturePool::addSegmentToFront(), featurePool, LexicalTree::findBestToken(), FeaturePool::getCurSegmentLen(), LexicalTree::grammarStart, WLRType::isSil, TokenType::likelihood, WLRType::lmHistory, LexicalNode::modelID, LexicalNode::parallel, TokenType::path, WLRType::previous, WLRType::timeStamp, and LexicalNode::tokenSeq.

Referenced by segmentFeaturePool().

Here is the call graph for this function:

int Segmenter::getNumberOfModels ( ) const [inline]

References numberOfClusters.

float Segmenter::segmentFeaturePool	(	int	inputSegID,
		int	inputLabelID,
		int	outputSegID
	)

Given the current feature pool and input segmentation and label to use, a new output segmentation is created. If the inputSegID is smaller than zero, the entire featurePool is used as one segment. If inputLabelID is smaller than zero, all labels in the input segmentation are used.

References FeaturePool::claimSegmentationID(), createSegments(), featurePool, FeaturePool::freeSegmentationID(), FeaturePool::getCurSegmentStart(), FeaturePool::getFirstVectorFirstSegmentList(), FeaturePool::getFirstVectorNextSegmentList(), FeaturePool::getNextVectorList(), FeaturePool::getVectorList(), LexicalTree::initialiseTree(), LexicalTree::processVector(), FeaturePool::resetSegmentation(), FeaturePool::segmentationIntersection(), and FeaturePool::startTimeStamp().