zcpa1.cpp

这个程序是基于zcpa语音特征提取方法的

// CZCPA.cpp: implementation of the CZCPA class.
//
//////////////////////////////////////////////////////////////////////

#include "zcpa.h"
#include <math.h>
#include <stdlib.h>

CZCPA::CZCPA()
{
	int i;

	iFreqDimension = 16;
	iNumFirTab = 100;
	
	
	piChannel = new int [iFreqDimension];
	piCenterFreq = new int [iFreqDimension];
	ppdFirCoeff = new double* [iNumFirTab];
	for( i=0 ; i<iNumFirTab ; i++ )
		ppdFirCoeff[i] = new double [iFreqDimension];
}

CZCPA::~CZCPA()
{
	delete [] piChannel;
	delete [] piCenterFreq;
	for( int i=0 ; i<iNumFirTab ; i++ )
		delete [] ppdFirCoeff[i];
	delete [] ppdFirCoeff;
}

void CZCPA::InitZCPA()
{

	int i,j;


	FILE *fp;

	fp = fopen(ZCPA_PARAMETER_FLOATING_FILE,"r");
	if( !fp )
	{
		fprintf(stderr,"ZCPA Floating Parameter Open Error");
		exit(1);
	}
	
	for( i=0 ; i<iFreqDimension ; i++ )
		fscanf(fp,"%d",&piChannel[i]);

	for( i=0 ; i<iFreqDimension ; i++ )
		fscanf(fp,"%d",&piCenterFreq[i]);


	for( i=0 ; i<iNumFirTab ; i++ )
		for( j=0 ; j<iFreqDimension ; j++ )
			fscanf(fp,"%lf",&ppdFirCoeff[i][j]);

	for( i=0 ; i<16 ; i++ )
		dBaseMem[i] = 0;
	for( i=0 ; i<7 ; i++ )
		dMem1[i] = 0;
	for( i=0 ; i<4 ; i++ )
		dMem2[i] = 0;
	for( i=0 ; i<2 ; i++ )
		dMem3[i] = 0;
	dMem4 = 0;
}

inline int CZCPA::CalCrovPoint(double y1, double y2)
{
	return( (int)(-64.*y1/(-y1+y2)) );
}

inline double CZCPA::Nonlinear(double peak)
{
	return( log10(1.0+20.0*peak)) ;
}

void CZCPA::TimeNormalize(double **x_in, int len_in, double *x_out, int len_out)
{
	int i,j,k;
	double *cum_dist, rate0, rate1, deno, delta_dist, dif, tmp;

	len_in = len_in-1;
	cum_dist = new double [len_in];

	for(i=0,cum_dist[0]=0.,tmp=0. ; i<len_in-1 ; i++){
		tmp+=GetDist(x_in[i+1],x_in[i+2],iFreqDimension);
		cum_dist[i+1]=tmp;
	}
	delta_dist=cum_dist[len_in-1]/(double)(len_out-1);

	for(i=0;i<iFreqDimension;i++) x_out[i]=x_in[1][i];

	for(i=1; i<len_out-1;i++){
		dif=delta_dist*(double)i;
		for(j=0;j<len_in-1;j++)
			if(dif>=cum_dist[j]&&dif<cum_dist[j+1]){
				deno=cum_dist[j+1]-cum_dist[j];
				rate0=(dif-cum_dist[j])/deno;
				rate1=(cum_dist[j+1]-dif)/deno;

				for(k=0;k<iFreqDimension;k++)
					x_out[i*iFreqDimension+k]=(double)(rate0*x_in[j+2][k]+rate1*x_in[j+1][k]);
			}
	}
  
	for(k=0;k<iFreqDimension;k++)
			x_out[(len_out-1)*iFreqDimension+k]=x_in[len_in+1-1][k];
	delete [] cum_dist;	cum_dist = NULL;

}

void CZCPA::AmpNormalize(double *x_in, int iLen)
{
	int i;

	double dSum;

	for( i=0, dSum=0 ; i<iLen ; i++ )
	{
		dSum += x_in[i];
	}

	dSum/=(double)iLen;

	for( i=0 ; i<iLen ; i++ )
	{
		x_in[i]/=dSum;
	}
}

double CZCPA::GetDist(double *x, double *y, int dim)
{
	int i;
	double distance;

	for(i=0, distance=0.;i<dim;i++)
		if(x[i]>y[i]) distance+=(x[i]-y[i]);
		else distance+=(y[i]-x[i]);

	return(distance);
}

void CZCPA::CalZCPA(short *input, int iSpeechLen, double *feat)
{
	int i,j,k;
	double **ppdFilterOut;
	int  nFrame;
	int inCLK;
	double max, peak;
	int intv;
	int dist, freq_bin;
	int zcpa_count;
	double **zcpa;
	short int *speech;

	nFrame = (iSpeechLen+99)/110;
	
	for( i=0 ; i<iSpeechLen ; i++ )
		input[i] = input[i]>>4;

	ppdFilterOut = new double * [iSpeechLen+99];
	for( i=0 ; i<iSpeechLen+99 ; i++)
		ppdFilterOut[i] = new double [iFreqDimension];

	for( i=0 ; i<iSpeechLen+99 ; i++)
		for( j=0 ; j<iFreqDimension; j++)
			ppdFilterOut[i][j] = 0.;

	for( i=0; i<iFreqDimension; i++)
	{
		for( j=0; j<iSpeechLen+99; j++)
		{
			for( k=0; k<iNumFirTab; k++)
				if( (j-k) >= 0 && (j-k) < iSpeechLen )
					ppdFilterOut[j][i] += (double)input[j-k]*ppdFirCoeff[k][i];
		}
	}

	zcpa = new double * [nFrame+5];
	for( i=0 ; i<nFrame+5 ; i++)
		zcpa[i] = new double [iFreqDimension];

	for( i=0; i<nFrame+5; i++)
		for( j=0; j<iFreqDimension; j++)
			zcpa[i][j] = 0.;
	
	// convolution
	// extract ZCPA
	for( i=0; i<iFreqDimension; i++)
	{
		inCLK = 0;
		max = 0;
		intv = 64;
		zcpa_count = 0;

		for( j=0; j<iFreqDimension; j++)
		{
			dBaseMem[j] = 0;
			if( j<7)
				dMem1[j] = 0;
			if( j<4)
				dMem2[j] = 0;
			if( j<2)
				dMem3[j] = 0;
			if( j<1)
				dMem4 = 0;
		}

		for( j=0; j<iSpeechLen+99; j++)
		{

			// zero-crossing detection
			if( j != 0 && (ppdFilterOut[j][i] >= 0 && ppdFilterOut[j-1][i] < 0) )
			{
				dist = CalCrovPoint( ppdFilterOut[j-1][i], ppdFilterOut[j][i]);
				intv += dist;
				peak = Nonlinear(max);
				max = 0.;

				if( intv <= 4525 && intv > 174 )
				{
					for( k=0; k<iFreqDimension; k++)
					{
						if( k == 15 )
						{
							if( intv >= piCenterFreq[k] )
								freq_bin = k;	
						}
						else if( intv >= piCenterFreq[k] && intv < piCenterFreq[k+1] )
							freq_bin = k;
					}

					if( inCLK <= piChannel[i] )
					{
						dBaseMem[freq_bin] += peak;
					}
					if( i > 8 )
					{
						dMem1[freq_bin-9] += peak;
					}

				}

				intv = -dist;
			}

			if( ppdFilterOut[j][i] >= 0)
				if( ppdFilterOut[j][i] > max )
					max = ppdFilterOut[j][i];

			if( intv < 6144 )
				intv += 64;

	
			if( inCLK == 109 )
				inCLK = 0;
			else
				inCLK++;

			if( inCLK == 0 )
			{

				for( k=0; k<iFreqDimension; k++)
					zcpa[zcpa_count][k] += dBaseMem[k];

				for( k=0; k<iFreqDimension; k++)
				{
					if( k > 14 )
					{
						dBaseMem[k] = dMem4 + dMem3[k-14] + dMem2[k-12] + dMem1[k-9];
						dMem4 = dMem3[1];
						dMem3[1] = dMem2[3];
						dMem2[3] = dMem1[6];
						dMem1[6] = 0;
					}
					else if( k > 13 )
					{
						dBaseMem[k] = dMem3[k-14] + dMem2[k-12] + dMem1[k-9];
						dMem3[k-14] = dMem2[k-12];
						dMem2[k-12] = dMem1[k-9];
						dMem1[k-9] = 0;
					}
					else if( k > 11 )
					{
						dBaseMem[k] = dMem2[k-12] + dMem1[k-9];
						dMem2[k-12] = dMem1[k-9];
						dMem1[k-9] = 0;
					}
					else if( k > 8 )
					{
						dBaseMem[k] = dMem1[k-9];
						dMem1[k-9] = 0;
					}
					else 
						dBaseMem[k] = 0;
				}
				zcpa_count++;
			}
		}
	}	


	for( i=0 ; i<iSpeechLen+99 ; i++ )
		delete [] ppdFilterOut[i];
	delete [] ppdFilterOut;	ppdFilterOut = NULL;

	TimeNormalize( zcpa, zcpa_count, feat, 64);
	AmpNormalize( feat, 1024);


	for( i=0 ; i<nFrame+5 ; i++ )
		delete [] zcpa[i];
	delete [] zcpa;	zcpa = NULL;

}