Train_Segmenter Class Reference

This class can 'train' new speaker clusters. One could call determining the clusters training because the phone (SIL) training procedures are used. But in fact, training is done on the target data (not on a training set), so it is not really training but processing. See our Spring 2006 NIST Rich Transcription Speaker Diarization paper for more information. This is ment to be the most simple form of the HMM-based merging diarization method. Overloaded classes may be created for testing new algorithms. More...

Inheritance diagram for Train_Segmenter:

Inheritance graph
[legend]

List of all members.


Public Member Functions

 Train_Segmenter (FeaturePool *fp, int sID_train, int sID_decode, const char *name, int nrMerges, int minClusters, int maxClusters, bool widen, const char *nonMergeableModels)
virtual ~Train_Segmenter ()
void train (int maxClusters, char *label, char *tempStr, char *feaPosteriors, int segID_cheat=-1)
int getSegmentation () const
void storeClusters (FILE *outFile, char *outputName)
void loadClusters (FILE *inFile)

Protected Member Functions

double trainIteration (int nrIterations)
double trainClusters (int nrG=-1)
bool mergeClusters (int maxClusters, bool smallestWins, bool actuallyMerge)
void createInitialModels (int segAmount, char *outputName=NULL)
void createOverlapTree (int min)
virtual void startMergeIteration ()
virtual void getMergeModelScore (int model1, int model2, double *mergeScore)
virtual bool proceedMerge (int model1, int model2, int method)
virtual void mergeModels (int model1, int model2)
virtual void trainModel (int model, int nrG, double *trainPRes)
void getOverlap (FILE *file)
void writePosteriors (char *fileName)

Protected Attributes

char outputFileName [2000]
char label [200]
int sadID_train
int sadID_decode
int segID
int segID_decode
int segID_cheat
TrainPhoneModel ** trainCluster
TrainPhoneModel ** scoreCluster
int nrScoreClusters
Vector ** distanceVect
bool * mergeable
int nrNonMergeModels
double * clusterScore
int * clusterSize
int cheatSpk1
int cheatSpk2
int numberOfMerges
int * compareClusters
PhoneModel * vtlnModel
int fastMerge
bool prepareForASR
int maxDataPoints
bool detectOverlap
int nrMergeIterations
double bic_meanoffset [10]
TrainPhoneModel * trcl

Detailed Description

This class can 'train' new speaker clusters. One could call determining the clusters training because the phone (SIL) training procedures are used. But in fact, training is done on the target data (not on a training set), so it is not really training but processing. See our Spring 2006 NIST Rich Transcription Speaker Diarization paper for more information. This is ment to be the most simple form of the HMM-based merging diarization method. Overloaded classes may be created for testing new algorithms.

Constructor & Destructor Documentation

Train_Segmenter::Train_Segmenter ( FeaturePool *  fp,
int  sID_train,
int  sID_decode,
const char *  name,
int  nrMerges,
int  minClusters,
int  maxClusters,
bool  widen,
const char *  nonMergeableModels 
)

This constructor will initialise a LexicalTree topology with (hopefully) too many clusters.

References bic_meanoffset, FeaturePool::claimSegmentationID(), cluster_thread, cluster_threadMayStart, cluster_threadRunning, clusterScore, clusterSize, compareClusters, createOverlapTree(), detectOverlap, distanceVect, TrainPhoneModel::doNotuseBordersForTraining(), fastMerge, Segmenter::featurePool, WriteFileLittleBigEndian::freadEndianSafe(), FeaturePool::getClusterLength(), FeaturePool::getStreamInfo(), FeaturePool::getVectorSize(), Thread_Train_Data_Cluster::id, LexicalTree::languageModel, maxDataPoints, mergeable, nrMergeIterations, nrNonMergeModels, Segmenter::numberOfClusters, numberOfMerges, LexicalTree::numberOfPhones, LexicalTree::numberOfWords, LexicalTree::overwriteWeightPars(), LexicalTree::phoneModels, prepareForASR, TrainPhoneModel::readModel(), sadID_decode, sadID_train, segID, segID_decode, Segmenter::setFeaturePool(), TrainPhoneModel::setTrainingData(), thread_train_cluster(), trainCluster, trcl, and LexicalTree::vocabulary.

Here is the call graph for this function:

Train_Segmenter::~Train_Segmenter (  )  [virtual]

The destructor will delete all claimed memory (the clusters)

References clusterScore, compareClusters, Segmenter::featurePool, FeaturePool::freeSegmentationID(), LexicalTree::numberOfPhones, LexicalTree::phoneModels, segID, and trainCluster.

Here is the call graph for this function:

Member Function Documentation

void Train_Segmenter::createInitialModels ( int  segAmount,
char *  outputName = NULL 
) [protected]

We have done training, let's start merging!

References Segmenter::featurePool, FeaturePool::getClusterLength(), LexicalTree::numberOfWords, sadID_train, segID, trainClusters(), and trainIteration().

Referenced by train().

Here is the call graph for this function:

void Train_Segmenter::createOverlapTree ( int  min  )  [protected]

References Segmenter::createLexicalTree(), LexicalTree::numberOfPhones, LexicalTree::numberOfWords, and trainCluster.

Referenced by getOverlap(), loadClusters(), train(), Train_Segmenter(), trainIteration(), and trainModel().

Here is the call graph for this function:

void Train_Segmenter::getMergeModelScore ( int  model1,
int  model2,
double *  mergeScore 
) [protected, virtual]

Merges two single models and calculates the total score..

References Segmenter::featurePool, FeaturePool::getStreamInfo(), TrainPhoneModel::getTrainSilP(), FeaturePoolInfo::numberOfChannels, FeaturePool::resetStreamWeights(), segID, FeaturePool::setStreamWeights(), TrainPhoneModel::setTrainingData(), TrainPhoneModel::train(), trainCluster, and weights.

Referenced by mergeClusters().

Here is the call graph for this function:

void Train_Segmenter::getOverlap ( FILE *  file  )  [protected]

Todo:
Docs

References FeaturePool::addSegment(), FeaturePool::claimSegmentationID(), clusterScore, clusterSize, Segmenter::createLexicalTree(), createOverlapTree(), SegmentationAdmin::curSeg, TrainPhoneModel::doNotuseBordersForTraining(), Segmenter::featurePool, SegmentationList::firstFrame, FeaturePool::freeSegmentationID(), FeaturePool::getClusterLength(), FeaturePool::getCurSegmentLen(), FeaturePool::getCurSegmentStart(), FeaturePool::getFirstVectorFirstSegment(), FeaturePool::getFirstVectorNextSegment(), PhoneModel::getLogPDFProbability(), FeaturePool::getNextVector(), FeaturePool::getSegmentID(), FeaturePool::getStreamInfo(), TrainPhoneModel::getTrainSilP(), FeaturePool::getVectorSize(), SegmentationList::ID, label, SegmentationList::lastFrame, mergeClusters(), SegmentationList::next, LexicalTree::numberOfWords, outputFileName, SegmentationAdmin::prevSeg, sadID_decode, sadID_train, segID, FeaturePool::segmentationCreateOverlap(), Segmenter::segmentFeaturePool(), TrainPhoneModel::setTrainingData(), TrainPhoneModel::train(), trainCluster, trainClusters(), LexicalTree::vocabulary, PhoneModel::writeModel(), and FeaturePool::writeRTTM().

Referenced by train().

Here is the call graph for this function:

int Train_Segmenter::getSegmentation (  )  const [inline]

References segID_decode.

void Train_Segmenter::loadClusters ( FILE *  inFile  ) 

This method will load the trained clusters from inFile.

References FeaturePool::claimSegmentationID(), createOverlapTree(), TrainPhoneModel::doNotuseBordersForTraining(), Segmenter::featurePool, WriteFileLittleBigEndian::freadEndianSafe(), FeaturePool::freeSegmentationID(), FeaturePool::getStreamInfo(), FeaturePool::getVectorSize(), Segmenter::numberOfClusters, LexicalTree::numberOfWords, TrainPhoneModel::readModel(), sadID_train, segID, segID_cheat, segID_decode, FeaturePool::segmentationIntersection(), Segmenter::segmentFeaturePool(), TrainPhoneModel::setTrainingData(), and trainCluster.

Referenced by train().

Here is the call graph for this function:

bool Train_Segmenter::mergeClusters ( int  maxClusters,
bool  smallestWins,
bool  actuallyMerge 
) [protected]

This method reduces the network with a single state (cluster).

References bic_meanoffset, cheatSpk1, cheatSpk2, fastMerge, Segmenter::featurePool, FeaturePool::getClusterLength(), Gaussian::getMean(), getMergeModelScore(), FeaturePool::getStreamInfo(), FeaturePool::getStreamWeight(), Vector::getValue(), Gaussian::getVariance(), globalIt, label, mergeable, mergeModels(), nrMergeIterations, FeaturePoolInfo::numberOfChannels, Segmenter::numberOfClusters, numberOfMerges, LexicalTree::numberOfWords, outputFileName, segID, FeaturePool::setStreamWeights(), Vector::setValue(), startMergeIteration(), Gaussian::train(), trainCluster, and Gaussian::trainFinish().

Referenced by getOverlap(), and train().

Here is the call graph for this function:

void Train_Segmenter::mergeModels ( int  model1,
int  model2 
) [protected, virtual]

Merges two single models (over loaded by Adapt_Segmenter)

Reimplemented in Adapt_Segmenter.

References FeaturePool::claimSegmentationID(), Segmenter::featurePool, FeaturePool::freeSegmentationID(), globalIt, label, FeaturePool::mergeLabels(), Segmenter::numberOfClusters, outputFileName, segID, segID_cheat, segID_decode, FeaturePool::segmentationIntersection(), TrainPhoneModel::setTrainingData(), TrainPhoneModel::train(), and trainCluster.

Referenced by mergeClusters().

Here is the call graph for this function:

bool Train_Segmenter::proceedMerge ( int  model1,
int  model2,
int  method 
) [protected, virtual]

We have decided on two merge candidates. Now determine if we should actually do it, or just stop.

Reimplemented in Adapt_Segmenter.

void Train_Segmenter::startMergeIteration (  )  [protected, virtual]

Before we check which model to merge, we do some preparations in this method.

Reimplemented in Adapt_Segmenter.

References compareClusters, and LexicalTree::numberOfWords.

Referenced by mergeClusters().

void Train_Segmenter::storeClusters ( FILE *  outFile,
char *  outputName 
)

This method will store the trained clusters to outFile.

References FeaturePool::claimSegmentationID(), Segmenter::featurePool, FeaturePool::getClusterLength(), FeaturePool::getNumberOfSpeakers(), FeaturePool::getSpeakerID(), FeaturePool::getVectorSize(), Segmenter::numberOfClusters, FeaturePool::readRTTM(), sadID_decode, trainCluster, and PhoneModel::writeModel().

Here is the call graph for this function:

void Train_Segmenter::train ( int  maxClusters,
char *  label,
char *  tempStr,
char *  feaPosteriors,
int  segID_cheat = -1 
)

References FeaturePool::addSegment(), FeaturePool::claimSegmentationID(), cluster_threadRunning, condition_cluster_mutexDone, condition_cluster_threadDone, createInitialModels(), Segmenter::createLexicalTree(), createOverlapTree(), detectOverlap, Segmenter::featurePool, FeaturePool::freeSegmentationID(), FeaturePool::getClusterLength(), FeaturePool::getCurSegmentLen(), FeaturePool::getCurSegmentStart(), FeaturePool::getFirstVectorFirstSegment(), FeaturePool::getFirstVectorNextSegment(), getOverlap(), FeaturePool::getStreamInfo(), globalIt, LexicalTree::initialiseSystem(), label, loadClusters(), loadedOverlapAM, mergeClusters(), nrNonMergeModels, FeaturePoolInfo::numberOfChannels, Segmenter::numberOfClusters, LexicalTree::numberOfWords, outputFileName, prepareForASR, FeaturePool::resetSegmentation(), sadID_decode, sadID_train, segID, segID_cheat, segID_decode, SEGMENT_INITIAL_DATA_SEGMENTS, FeaturePool::segmentationCreateOverlap(), FeaturePool::segmentationWiden(), Segmenter::segmentFeaturePool(), trainCluster, trainIteration(), trainModel(), FeaturePoolInfo::weight, writePosteriors(), and FeaturePool::writeRTTM().

Referenced by Shout_Cluster::Shout_Cluster().

Here is the call graph for this function:

double Train_Segmenter::trainClusters ( int  nrG = -1  )  [protected]

The training iterations are performed by this method.

References cluster_threadRunning, clusterScore, condition_cluster_mutexDone, condition_cluster_threadDone, TrainPhoneModel::maxNrOfGaussians(), mergeable, LexicalTree::numberOfWords, trainCluster, and trainModel().

Referenced by createInitialModels(), getOverlap(), and trainIteration().

Here is the call graph for this function:

double Train_Segmenter::trainIteration ( int  nrIterations  )  [protected]

The main method of a train iteration:

References LanguageModel_Segmenter::addPenalty(), FeaturePool::claimSegmentationID(), createOverlapTree(), Segmenter::featurePool, LanguageModel_Segmenter::finishModel(), FeaturePool::freeSegmentationID(), FeaturePool::getFirstVectorFirstSegment(), FeaturePool::getFirstVectorNextSegment(), FeaturePool::getSegmentID(), LexicalTree::languageModel, Segmenter::numberOfClusters, LexicalTree::numberOfWords, sadID_train, segID, segID_cheat, segID_decode, FeaturePool::segmentationIntersection(), Segmenter::segmentFeaturePool(), TrainPhoneModel::setTrainingData(), trainCluster, trainClusters(), and LanguageModel_Segmenter::transferTo().

Referenced by createInitialModels(), and train().

Here is the call graph for this function:

void Train_Segmenter::trainModel ( int  model,
int  nrG,
double *  trainPRes 
) [protected, virtual]

Trains a single model with nrG gaussians (over loaded by Adapt_Segmenter)

References cluster_threadMayStart, cluster_threadRunning, condition_cluster_mutexDone, condition_cluster_mutexStart, condition_cluster_threadDone, condition_cluster_threadStart, createOverlapTree(), Segmenter::discrTrain, Segmenter::discrTrainMask, FeaturePool::doSegmentsExist(), Segmenter::featurePool, TrainPhoneModel::getTrainSilP(), mergeable, Thread_Train_Data_Cluster::model, Thread_Train_Data_Cluster::nr, Thread_Train_Data_Cluster::segid, segID, TrainPhoneModel::setTrainingData(), TrainPhoneModel::train(), trainCluster, and Thread_Train_Data_Cluster::trainPRes.

Referenced by train(), and trainClusters().

Here is the call graph for this function:

void Train_Segmenter::writePosteriors ( char *  fileName  )  [protected]

Todo:
Docs

References Segmenter::featurePool, PhoneModel::getPDFProbability(), FeaturePool::getPoolLength(), FeaturePool::getVector(), LexicalTree::numberOfWords, and trainCluster.

Referenced by train().

Here is the call graph for this function:

Member Data Documentation

double Train_Segmenter::bic_meanoffset[10] [protected]

Referenced by mergeClusters(), and Train_Segmenter().

int Train_Segmenter::cheatSpk1 [protected]

Referenced by mergeClusters().

int Train_Segmenter::cheatSpk2 [protected]

Referenced by mergeClusters().

double* Train_Segmenter::clusterScore [protected]

Referenced by getOverlap(), Train_Segmenter(), trainClusters(), and ~Train_Segmenter().

int* Train_Segmenter::clusterSize [protected]

Referenced by getOverlap(), and Train_Segmenter().

int* Train_Segmenter::compareClusters [protected]

Referenced by Adapt_Segmenter::Adapt_Segmenter(), Adapt_Segmenter::proceedMerge(), startMergeIteration(), Train_Segmenter(), and ~Train_Segmenter().

bool Train_Segmenter::detectOverlap [protected]

Referenced by train(), and Train_Segmenter().

Vector** Train_Segmenter::distanceVect [protected]

Referenced by Train_Segmenter().

int Train_Segmenter::fastMerge [protected]

Referenced by mergeClusters(), and Train_Segmenter().

char Train_Segmenter::label[200] [protected]

Reimplemented in Adapt_Segmenter.

Referenced by getOverlap(), mergeClusters(), mergeModels(), and train().

int Train_Segmenter::maxDataPoints [protected]

Referenced by Train_Segmenter().

bool* Train_Segmenter::mergeable [protected]

Referenced by mergeClusters(), Train_Segmenter(), trainClusters(), and trainModel().

int Train_Segmenter::nrMergeIterations [protected]

Referenced by mergeClusters(), and Train_Segmenter().

int Train_Segmenter::nrNonMergeModels [protected]

Referenced by train(), and Train_Segmenter().

int Train_Segmenter::nrScoreClusters [protected]

int Train_Segmenter::numberOfMerges [protected]

Referenced by mergeClusters(), and Train_Segmenter().

char Train_Segmenter::outputFileName[2000] [protected]

Referenced by getOverlap(), mergeClusters(), mergeModels(), and train().

bool Train_Segmenter::prepareForASR [protected]

Referenced by train(), and Train_Segmenter().

int Train_Segmenter::sadID_decode [protected]

Referenced by getOverlap(), storeClusters(), train(), and Train_Segmenter().

int Train_Segmenter::sadID_train [protected]

Referenced by createInitialModels(), getOverlap(), loadClusters(), Adapt_Segmenter::proceedMerge(), train(), Train_Segmenter(), and trainIteration().

TrainPhoneModel** Train_Segmenter::scoreCluster [protected]

int Train_Segmenter::segID [protected]

Referenced by createInitialModels(), getMergeModelScore(), getOverlap(), loadClusters(), mergeClusters(), mergeModels(), Adapt_Segmenter::proceedMerge(), train(), Train_Segmenter(), trainIteration(), trainModel(), and ~Train_Segmenter().

int Train_Segmenter::segID_cheat [protected]

Referenced by loadClusters(), mergeModels(), train(), and trainIteration().

int Train_Segmenter::segID_decode [protected]

Referenced by getSegmentation(), loadClusters(), mergeModels(), train(), Train_Segmenter(), and trainIteration().

TrainPhoneModel** Train_Segmenter::trainCluster [protected]

Referenced by createOverlapTree(), getMergeModelScore(), getOverlap(), loadClusters(), mergeClusters(), mergeModels(), Adapt_Segmenter::proceedMerge(), storeClusters(), train(), Train_Segmenter(), trainClusters(), trainIteration(), trainModel(), writePosteriors(), and ~Train_Segmenter().

TrainPhoneModel* Train_Segmenter::trcl [protected]

Referenced by Train_Segmenter().

PhoneModel* Train_Segmenter::vtlnModel [protected]