c@243: /*
c@243:  *  ClusterMeltSegmenter.cpp
c@243:  *  soundbite
c@243:  *
c@243:  *  Created by Mark Levy on 23/03/2006.
c@243:  *  Copyright 2006 Centre for Digital Music, Queen Mary, University of London. All rights reserved.
c@243:  *
c@243:  */
c@243: 
c@243: #include <cfloat>
c@243: #include <cmath>
c@243: 
c@243: #include "ClusterMeltSegmenter.h"
c@243: #include "lib_constQ.h"
c@243: #include "cluster_segmenter.h"
c@243: #include "segment.h"
c@243: 
c@243: ClusterMeltSegmenter::ClusterMeltSegmenter(ClusterMeltSegmenterParams params) : window(NULL),
c@243: constq(NULL),
c@243: featureType(params.featureType),
c@243: windowSize(params.windowSize),
c@243: hopSize(params.hopSize),
c@243: fmin(params.fmin),
c@243: fmax(params.fmax),
c@243: nbins(params.nbins),
c@243: ncomponents(params.ncomponents),	// NB currently not passed - no. of PCA components is set in cluser_segmenter.c
c@243: nHMMStates(params.nHMMStates),
c@243: nclusters(params.nclusters),
c@243: histogramLength(params.histogramLength),
c@243: neighbourhoodLimit(params.neighbourhoodLimit)
c@243: {
c@243: }
c@243: 
c@243: void ClusterMeltSegmenter::initialise(int fs)
c@243: {
c@243: 	samplerate = fs;
c@243: 	if (featureType != FEATURE_TYPE_UNKNOWN)
c@243: 	{
c@243: 		ncoeff = static_cast<int>(ceil(nbins * (log(fmax / static_cast<double>(fmin))) / log(2.0)));
c@243: 		constq = init_constQ(fmin, fmax, nbins, samplerate, ncoeff);	
c@243: 	}
c@243: }
c@243: 
c@243: ClusterMeltSegmenter::~ClusterMeltSegmenter() 
c@243: {
c@243: 	delete [] window;
c@243: 	if (constq)
c@243: 		close_constQ(constq);
c@243: }
c@243: 
c@243: void ClusterMeltSegmenter::extractFeatures(double* samples, int nsamples)
c@243: {
c@243: 	// create a new window if needed
c@243: 	if (!window || nsamples != windowLength)
c@243: 	{
c@243: 		if (window)
c@243: 			delete [] window;
c@243: 		window = hamming_p(nsamples);
c@243: 		windowLength = nsamples;
c@243: 	}	
c@243: 	
c@243: 	// copy the samples before windowing in case we need them for something else
c@243: 	double* frame = new double[nsamples];
c@243: 	for (int i = 0; i < nsamples; i++)
c@243: 		frame[i] = samples[i] * window[i];
c@243: 	
c@243: 	// extract const-Q
c@243: 	do_constQ(constq, frame, nsamples);
c@243: 	int ncq = constq->ncoeff;
c@243: 	
c@243: 	delete [] frame;
c@243: 	
c@243: 	if (ncq == ncoeff)		// else feature extraction failed
c@243: 	{
c@243: 		vector<double> cq(ncq);
c@243: 		for (int i = 0; i < ncq; i++)
c@243: 			cq[i] = constq->absconstQtransform[i];
c@243: 		features.push_back(cq);
c@243: 	}
c@243: }
c@243: 
c@243: void ClusterMeltSegmenter::segment(int m)
c@243: {
c@243: 	nclusters = m;
c@243: 	segment();
c@243: }
c@243: 
c@243: void ClusterMeltSegmenter::setFeatures(const vector<vector<double> >& f)
c@243: {
c@243: 	features = f;
c@243: 	featureType = FEATURE_TYPE_UNKNOWN;
c@243: }
c@243: 
c@243: void ClusterMeltSegmenter::segment()
c@243: {
c@243: 	if (constq)
c@243: 	{
c@243: 		close_constQ(constq);		// finished extracting features
c@243: 		constq = NULL;
c@243: 	}
c@243: 	
c@243: 	// for now copy the features to a native array and use the existing C segmenter...
c@243: 	double** arrFeatures = new double*[features.size()];	
c@243: 	for (int i = 0; i < features.size(); i++)
c@243: 	{
c@243: 		if (featureType == FEATURE_TYPE_UNKNOWN)
c@243: 			arrFeatures[i] = new double[features[0].size()];
c@243: 		else
c@243: 			arrFeatures[i] = new double[ncoeff+1];	// allow space for the normalised envelope
c@243: 		for (int j = 0; j < ncoeff; j++)
c@243: 			arrFeatures[i][j] = features[i][j];	
c@243: 	}
c@243: 	
c@243: 	q = new int[features.size()];
c@243: 	
c@243: 	if (featureType == FEATURE_TYPE_UNKNOWN)
c@243: 		cluster_segment(q, arrFeatures, features.size(), features[0].size(), nHMMStates, histogramLength, 
c@243: 						nclusters, neighbourhoodLimit);
c@243: 	else
c@243: 		constq_segment(q, arrFeatures, features.size(), nbins, ncoeff, featureType, 
c@243: 					nHMMStates, histogramLength, nclusters, neighbourhoodLimit);
c@243: 	
c@243: 	// convert the cluster assignment sequence to a segmentation
c@243: 	makeSegmentation(q, features.size());		
c@243: 	
c@243: 	// de-allocate arrays
c@243: 	delete [] q;
c@243: 	for (int i = 0; i < features.size(); i++)
c@243: 		delete [] arrFeatures[i];
c@243: 	delete [] arrFeatures;
c@243: 	
c@243: 	// clear the features
c@243: 	clear();
c@243: }
c@243: 
c@243: void ClusterMeltSegmenter::makeSegmentation(int* q, int len)
c@243: {
c@243: 	segmentation.segments.clear();
c@243: 	segmentation.nsegtypes = nclusters;
c@243: 	segmentation.samplerate = samplerate;
c@243: 	
c@243: 	Segment segment;
c@243: 	segment.start = 0;
c@243: 	segment.type = q[0];
c@243: 	
c@243: 	for (int i = 1; i < len; i++)
c@243: 	{
c@243: 		if (q[i] != q[i-1])
c@243: 		{
c@243: 			segment.end = i * getHopsize();
c@243: 			segmentation.segments.push_back(segment);
c@243: 			segment.type = q[i];
c@243: 			segment.start = segment.end;
c@243: 		}
c@243: 	}
c@243: 	segment.end = len * getHopsize();
c@243: 	segmentation.segments.push_back(segment);
c@243: }
c@243: 
c@243: /*
c@243: void ClusterMeltSegmenter::mpeg7ConstQ()
c@243: {
c@243: 	// convert to dB scale
c@243: 	for (int i = 0; i < features.size(); i++)
c@243: 		for (int j = 0; j < ncoeff; j++)
c@243: 			features[i][j] = 10.0 * log10(features[i][j] + DBL_EPSILON);
c@243: 	
c@243: 	// normalise features and add the norm at the end as an extra feature dimension
c@243: 	double maxnorm = 0;		// track the max of the norms
c@243: 	for (int i = 0; i < features.size(); i++)
c@243: 	{
c@243: 		double norm = 0;
c@243: 		for (int j = 0; j < ncoeff; j++)
c@243: 			norm += features[i][j] * features[i][j];
c@243: 		norm = sqrt(norm);
c@243: 		for (int j = 0; j < ncoeff; j++)
c@243: 			features[i][j] /= norm;
c@243: 		features[i].push_back(norm);
c@243: 		if (norm > maxnorm)
c@243: 			maxnorm = norm;
c@243: 	}
c@243: 	
c@243: 	// normalise the norms
c@243: 	for (int i = 0; i < features.size(); i++)
c@243: 		features[i][ncoeff] /= maxnorm;
c@243: }
c@243: */