cannam@18: /*
cannam@18:  *  ClusterMeltSegmenter.cpp
cannam@18:  *  soundbite
cannam@18:  *
cannam@18:  *  Created by Mark Levy on 23/03/2006.
cannam@18:  *  Copyright 2006 Centre for Digital Music, Queen Mary, University of London. All rights reserved.
cannam@18:  *
cannam@18:  */
cannam@18: 
cannam@18: #include <cfloat>
cannam@18: #include <cmath>
cannam@18: 
cannam@18: #include "ClusterMeltSegmenter.h"
cannam@18: #include "cluster_segmenter.h"
cannam@18: #include "segment.h"
cannam@18: 
cannam@20: #include "dsp/transforms/FFT.h"
cannam@20: 
cannam@18: ClusterMeltSegmenter::ClusterMeltSegmenter(ClusterMeltSegmenterParams params) : window(NULL),
cannam@18: constq(NULL),
cannam@18: featureType(params.featureType),
cannam@20: hopSize(params.hopSize),
cannam@18: windowSize(params.windowSize),
cannam@18: fmin(params.fmin),
cannam@18: fmax(params.fmax),
cannam@18: nbins(params.nbins),
cannam@18: ncomponents(params.ncomponents),	// NB currently not passed - no. of PCA components is set in cluser_segmenter.c
cannam@18: nHMMStates(params.nHMMStates),
cannam@18: nclusters(params.nclusters),
cannam@18: histogramLength(params.histogramLength),
cannam@18: neighbourhoodLimit(params.neighbourhoodLimit)
cannam@18: {
cannam@18: }
cannam@18: 
cannam@18: void ClusterMeltSegmenter::initialise(int fs)
cannam@18: {
cannam@18: 	samplerate = fs;
cannam@18: 	if (featureType != FEATURE_TYPE_UNKNOWN)
cannam@18: 	{
cannam@20: //!!!		ncoeff = static_cast<int>(ceil(nbins * (log(fmax / static_cast<double>(fmin))) / log(2.0)));
cannam@20: 		CQConfig config;
cannam@20: 		config.FS = samplerate;
cannam@20: 		config.min = fmin;
cannam@20: 		config.max = fmax;
cannam@20: 		config.BPO = nbins;
cannam@20: 		config.CQThresh = 0.0054;
cannam@20: 		constq = new ConstantQ(config);
cannam@20: //!!!		constq = init_constQ(fmin, fmax, nbins, samplerate, ncoeff);	
cannam@20: 		ncoeff = constq->getK();
cannam@18: 	}
cannam@18: }
cannam@18: 
cannam@18: ClusterMeltSegmenter::~ClusterMeltSegmenter() 
cannam@18: {
cannam@20: 	delete window;
cannam@20: 	delete constq;
cannam@20: //!!!	if (constq)
cannam@20: //		close_constQ(constq);
cannam@20: }
cannam@20: 
cannam@20: int
cannam@20: ClusterMeltSegmenter::getWindowsize()
cannam@20: {
cannam@20: 	if (featureType != FEATURE_TYPE_UNKNOWN) {
cannam@20: 		std::cerr << "rate = " << samplerate << ", fft length = " << constq->getfftlength() << ", fmin = " << fmin << ", fmax = " << fmax << ", nbins = " << nbins << ", K = " << constq->getK() << ", Q = " << constq->getQ() << std::endl;
cannam@20: 		return constq->getfftlength();
cannam@20: 	} else {
cannam@20: 		return static_cast<int>(windowSize * samplerate);
cannam@20: 	}
cannam@20: }
cannam@20: 
cannam@20: int
cannam@20: ClusterMeltSegmenter::getHopsize()
cannam@20: {
cannam@20: 	return static_cast<int>(hopSize * samplerate);
cannam@18: }
cannam@18: 
cannam@18: void ClusterMeltSegmenter::extractFeatures(double* samples, int nsamples)
cannam@18: {
cannam@18: 	// create a new window if needed
cannam@20: /*!!!
cannam@18: 	if (!window || nsamples != windowLength)
cannam@18: 	{
cannam@18: 		if (window)
cannam@18: 			delete [] window;
cannam@20: //		Window<double>(HammingWindow, nsamples).cut
cannam@20: //!!!		window = hamming_p(nsamples);
cannam@18: 		windowLength = nsamples;
cannam@18: 	}	
cannam@20: */
cannam@20: 	if (!window || window->getSize() != nsamples) {
cannam@20: 		delete window;
cannam@20: 		window = new Window<double>(HammingWindow, nsamples);
cannam@20: 	}
cannam@20: 
cannam@18: 	// copy the samples before windowing in case we need them for something else
cannam@18: 	double* frame = new double[nsamples];
cannam@20: //	for (int i = 0; i < nsamples; i++)
cannam@20: //		frame[i] = samples[i] * window[i];
cannam@20: 	window->cut(frame);
cannam@18: 	
cannam@20: 	std::cerr << "nsamples = " << nsamples << std::endl;
cannam@20: 
cannam@20: 	double *real = new double[nsamples];
cannam@20: 	double *imag = new double[nsamples];
cannam@20: 
cannam@20: 	FFT::process(nsamples, false, frame, 0, real, imag);
cannam@20: 
cannam@20: 	double *cqre = new double[ncoeff];
cannam@20: 	double *cqim = new double[ncoeff];
cannam@20: 
cannam@20: 	constq->process(real, imag, cqre, cqim);
cannam@20: 
cannam@18: 	// extract const-Q
cannam@20: //!!!	do_constQ(constq, frame, nsamples);
cannam@20: //	int ncq = constq->ncoeff;
cannam@20: 
cannam@20: 	delete [] frame;
cannam@20: 	delete [] real;
cannam@20: 	delete [] imag;
cannam@18: 	
cannam@20: //!!!	if (ncq == ncoeff)		// else feature extraction failed
cannam@20: //	{
cannam@20: //		vector<double> cq(ncq);
cannam@20: //		for (int i = 0; i < ncq; i++)
cannam@20: //			cq[i] = constq->absconstQtransform[i];
cannam@20: 		vector<double> cq(ncoeff);
cannam@20: 		for (int i = 0; i < ncoeff; ++i) {
cannam@20: 			cq[i] = sqrt(cqre[i] * cqre[i] + cqim[i] * cqim[i]);
cannam@20: 		}
cannam@18: 		features.push_back(cq);
cannam@20: //	}
cannam@20: 
cannam@20: 		delete[] cqre;
cannam@20: 		delete[] cqim;
cannam@18: }
cannam@18: 
cannam@18: void ClusterMeltSegmenter::segment(int m)
cannam@18: {
cannam@18: 	nclusters = m;
cannam@18: 	segment();
cannam@18: }
cannam@18: 
cannam@18: void ClusterMeltSegmenter::setFeatures(const vector<vector<double> >& f)
cannam@18: {
cannam@18: 	features = f;
cannam@18: 	featureType = FEATURE_TYPE_UNKNOWN;
cannam@18: }
cannam@18: 
cannam@18: void ClusterMeltSegmenter::segment()
cannam@18: {
cannam@18: 	if (constq)
cannam@18: 	{
cannam@20: //!!!		close_constQ(constq);		// finished extracting features
cannam@20: 		delete constq;
cannam@18: 		constq = NULL;
cannam@18: 	}
cannam@18: 	
cannam@18: 	// for now copy the features to a native array and use the existing C segmenter...
cannam@18: 	double** arrFeatures = new double*[features.size()];	
cannam@18: 	for (int i = 0; i < features.size(); i++)
cannam@18: 	{
cannam@18: 		if (featureType == FEATURE_TYPE_UNKNOWN)
cannam@18: 			arrFeatures[i] = new double[features[0].size()];
cannam@18: 		else
cannam@18: 			arrFeatures[i] = new double[ncoeff+1];	// allow space for the normalised envelope
cannam@18: 		for (int j = 0; j < ncoeff; j++)
cannam@18: 			arrFeatures[i][j] = features[i][j];	
cannam@18: 	}
cannam@18: 	
cannam@18: 	q = new int[features.size()];
cannam@18: 	
cannam@18: 	if (featureType == FEATURE_TYPE_UNKNOWN)
cannam@18: 		cluster_segment(q, arrFeatures, features.size(), features[0].size(), nHMMStates, histogramLength, 
cannam@18: 						nclusters, neighbourhoodLimit);
cannam@18: 	else
cannam@18: 		constq_segment(q, arrFeatures, features.size(), nbins, ncoeff, featureType, 
cannam@18: 					nHMMStates, histogramLength, nclusters, neighbourhoodLimit);
cannam@18: 	
cannam@18: 	// convert the cluster assignment sequence to a segmentation
cannam@18: 	makeSegmentation(q, features.size());		
cannam@18: 	
cannam@18: 	// de-allocate arrays
cannam@18: 	delete [] q;
cannam@18: 	for (int i = 0; i < features.size(); i++)
cannam@18: 		delete [] arrFeatures[i];
cannam@18: 	delete [] arrFeatures;
cannam@18: 	
cannam@18: 	// clear the features
cannam@18: 	clear();
cannam@18: }
cannam@18: 
cannam@18: void ClusterMeltSegmenter::makeSegmentation(int* q, int len)
cannam@18: {
cannam@18: 	segmentation.segments.clear();
cannam@18: 	segmentation.nsegtypes = nclusters;
cannam@18: 	segmentation.samplerate = samplerate;
cannam@18: 	
cannam@18: 	Segment segment;
cannam@18: 	segment.start = 0;
cannam@18: 	segment.type = q[0];
cannam@18: 	
cannam@18: 	for (int i = 1; i < len; i++)
cannam@18: 	{
cannam@18: 		if (q[i] != q[i-1])
cannam@18: 		{
cannam@18: 			segment.end = i * getHopsize();
cannam@18: 			segmentation.segments.push_back(segment);
cannam@18: 			segment.type = q[i];
cannam@18: 			segment.start = segment.end;
cannam@18: 		}
cannam@18: 	}
cannam@18: 	segment.end = len * getHopsize();
cannam@18: 	segmentation.segments.push_back(segment);
cannam@18: }
cannam@18: 
cannam@18: /*
cannam@18: void ClusterMeltSegmenter::mpeg7ConstQ()
cannam@18: {
cannam@18: 	// convert to dB scale
cannam@18: 	for (int i = 0; i < features.size(); i++)
cannam@18: 		for (int j = 0; j < ncoeff; j++)
cannam@18: 			features[i][j] = 10.0 * log10(features[i][j] + DBL_EPSILON);
cannam@18: 	
cannam@18: 	// normalise features and add the norm at the end as an extra feature dimension
cannam@18: 	double maxnorm = 0;		// track the max of the norms
cannam@18: 	for (int i = 0; i < features.size(); i++)
cannam@18: 	{
cannam@18: 		double norm = 0;
cannam@18: 		for (int j = 0; j < ncoeff; j++)
cannam@18: 			norm += features[i][j] * features[i][j];
cannam@18: 		norm = sqrt(norm);
cannam@18: 		for (int j = 0; j < ncoeff; j++)
cannam@18: 			features[i][j] /= norm;
cannam@18: 		features[i].push_back(norm);
cannam@18: 		if (norm > maxnorm)
cannam@18: 			maxnorm = norm;
cannam@18: 	}
cannam@18: 	
cannam@18: 	// normalise the norms
cannam@18: 	for (int i = 0; i < features.size(); i++)
cannam@18: 		features[i][ncoeff] /= maxnorm;
cannam@18: }
cannam@18: */