语音合成与识别

开发平台：
Visual C++

FE_vus.cpp：源码内容
							///////////////////////////////////////////////////////////////////////////////
// This is a part of the Feature program.
// Version: 1.0
// Date: February 22, 2003
// Programmer: Oh-Wook Kwon
// Copyright(c) 2003 Oh-Wook Kwon. All rights reserved. owkwon@ucsd.edu
///////////////////////////////////////////////////////////////////////////////
#include "StdAfx.h"
#include "FE_feature.h"
static const float EPD_NB_FRAME_THRESHOLD_LTE=(float)10;
static const float EPD_LAMBDA_LTE=(float)0.97;
static const float EPD_LAMBDA_ZCR=(float)0.97; // old=0.99
static const float EPD_SNR_THRESHOLD_UPD_LTE=(float)4;
static const float EPD_ZCR_THRESHOLD_UPD_LTE=(float)10;
static const float EPD_NOISE_ENERGY_FLOOR=(float)14.5;
static const float EPD_SPEECH_ENERGY_FLOOR=(float)37.5;
static const float lambdaLTEhigherE=(float)0.99;
static const int EPD_MIN_FRAME=10;
int Fe::vus_basic(short *sample, int sampleN, int frameSize, vector<EVusType>& vusA)
{
	int n,i;
	vector<float> energyA;
	vector<float> zcrA;
	vector<float> noiseEnA;
	vector<float> meanZcrA;
	float noiseEn=0;
	float meanZcr=0;
	int zcrFrameN=0;
#ifdef _DEBUG
#define TEST_FRAME_LEN 300
	float tmp1A[TEST_FRAME_LEN];
	float tmp2A[TEST_FRAME_LEN];
	EVusType vusTmpA[TEST_FRAME_LEN];
	float meanEnergyTmpA[TEST_FRAME_LEN];
	float meanZcrTmpA[TEST_FRAME_LEN];
#endif
	int shiftSize=GetShiftSize();
	int frameN=(int)((sampleN-(frameSize-shiftSize))/shiftSize);
	energyA.resize(frameN);		
	zcrA.resize(frameN);
	noiseEnA.resize(frameN);
	meanZcrA.resize(frameN);
	vusA.resize(frameN);
	float VUS_ENERGY_TH_VOICED=30;
	float VUS_ENERGY_TH_UNVOICED=60;
	float VUS_ENERGY_TH_START=20;
	float VUS_ENERGY_TH_INSPEECH=15; // old=6
	int VUS_ZCR_TH_MAX=50;
	int VUS_ZCR_TH_MIN=5;
	int VUS_VOICE_ZCR_TH_MIN=0;
	int VUS_ZCR_UPD_TH=50;
	vector<float> x(frameSize);
	float vus_energy_th=VUS_ENERGY_TH_START;
	for(n=0;n<frameN;n++){
		float lambdaLTE, lambdaZcr;
		if(n<EPD_NB_FRAME_THRESHOLD_LTE)
			lambdaLTE=1-1/(float)(n+1);
		else
			lambdaLTE=EPD_LAMBDA_LTE;
		if(zcrFrameN<EPD_NB_FRAME_THRESHOLD_LTE)
			lambdaZcr=1-1/(float)(zcrFrameN+1);
		else
			lambdaZcr=EPD_LAMBDA_ZCR;
		
		int begX=n*shiftSize;
		{
			/* DC offset removal, H(z)=(1-z^(-1))/(1-a*z^(-1)) */
			float a=1-1/(float)1024;
			x[0]=sample[begX];
			for(i=1;i<frameSize;i++){
				x[i]=(sample[begX+i]-sample[begX+i-1]+a*x[i-1]);
			}
			/* low-pass filter, x=filter([1 2 1]/4, [1], x); */
			vector<float> xorg=x;
			x[0]=(xorg[0]+2*xorg[0]+xorg[1])/4;
			for(i=1;i<frameSize-1;i++){
				x[i]=(xorg[i-1]+2*xorg[i]+xorg[i+1])/4;
			}
			x[frameSize-1]=(xorg[frameSize-2]+2*xorg[frameSize-1]+xorg[frameSize-1])/4;
		}
		/* Normalize energy to frame length */
		float sum=0;
		for(i=0;i<frameSize;i++){
			sum += (x[i]*x[i]);
		}
		energyA[n] = (0.5+10*LOG10(1+sum/frameSize)); /* dB scale */
		if(energyA[n]<EPD_NOISE_ENERGY_FLOOR) energyA[n]=EPD_NOISE_ENERGY_FLOOR;
		/* Estimate noise */
		float noiseLevel=0;
		if((energyA[n]-noiseEn)<EPD_SNR_THRESHOLD_UPD_LTE || n<EPD_MIN_FRAME || (energyA[n] < EPD_SPEECH_ENERGY_FLOOR)){
			if((energyA[n]<noiseEn) | (n<EPD_MIN_FRAME) | (energyA[n] < EPD_SPEECH_ENERGY_FLOOR)){
				noiseEn=noiseEn+(1-lambdaLTE)*(energyA[n]-noiseEn);
			}
			else{
				noiseEn=noiseEn+(1-lambdaLTEhigherE)*(energyA[n]-noiseEn);
			}
			/* noise level should be computed from the original noise energy */
			noiseLevel=(2*sqrt(exp(log(10)/10*(noiseEn-0.5))-1));
			if(noiseEn<EPD_NOISE_ENERGY_FLOOR) noiseEn=EPD_NOISE_ENERGY_FLOOR;
		}
		noiseEnA[n]=noiseEn;
		/* compute zero crossing rate (changed) */	
		float prev = x[0]-noiseLevel;
		int zcr=0;
		for(i=1;i<frameSize;i++){
			float val  = x[i]-noiseLevel;
			float ztmp=val*prev;
			if(ztmp<0) zcr = zcr+1;
			prev=val;
		}
		zcrA[n]=zcr;
		if(zcr>0 || zcrFrameN>0){
			zcrFrameN++;
		}
		if(zcrA[n]<VUS_ZCR_UPD_TH && ((zcrA[n]-meanZcr)<EPD_ZCR_THRESHOLD_UPD_LTE || zcrFrameN<EPD_MIN_FRAME)){
			meanZcr=meanZcr+(1-lambdaZcr)*(zcrA[n]-meanZcr);
		}
		meanZcrA[n]=meanZcr;
		float deltaEnergy=energyA[n]-noiseEn;
		float deltaZcr=zcrA[n]-meanZcr;
		if(deltaEnergy<VUS_ENERGY_TH_UNVOICED && deltaEnergy<vus_energy_th && deltaZcr<VUS_VOICE_ZCR_TH_MIN){
			vusA[n]=FRM_SILENCE;
			vus_energy_th=VUS_ENERGY_TH_START;
		}
		else if(deltaZcr>VUS_ZCR_TH_MAX){
			vus_energy_th=VUS_ENERGY_TH_START;
			vusA[n]=FRM_UNVOICED;
		}
		else if(deltaZcr>VUS_ZCR_TH_MIN && deltaEnergy<vus_energy_th){
			vus_energy_th=VUS_ENERGY_TH_START;
			vusA[n]=FRM_UNVOICED;
		}
		else if(zcr>VUS_ZCR_TH_MIN && deltaEnergy<vus_energy_th){
			vus_energy_th=VUS_ENERGY_TH_START;
			vusA[n]=FRM_UNVOICED;
		}
		else if(deltaEnergy>VUS_ENERGY_TH_VOICED){
			vus_energy_th=VUS_ENERGY_TH_INSPEECH;
			vusA[n]=FRM_VOICED;
		}
		else if(deltaEnergy>vus_energy_th && (deltaZcr>VUS_VOICE_ZCR_TH_MIN || zcr>VUS_ZCR_TH_MIN)){
			vus_energy_th=VUS_ENERGY_TH_INSPEECH;
			vusA[n]=FRM_VOICED;
		}
		else{
			vusA[n]=FRM_SILENCE;
			vus_energy_th=VUS_ENERGY_TH_START;
		}
		/* TRACE("%d  %f %f %f %f %dn",n,energyA[n],zcrA[n],deltaEnergy,deltaZcr,(int)vusA[n]); */
	}
#ifdef _DEBUG
	int k;
	for(k=0;k<my_min(TEST_FRAME_LEN,vusA.size());k++){
		tmp1A[k]=energyA[k];
		tmp2A[k]=zcrA[k];
		meanEnergyTmpA[k]=noiseEnA[k];
		meanZcrTmpA[k]=meanZcrA[k];
		vusTmpA[k]=vusA[k];
	}
#endif
	/* median filtering */
	vus_median_filter(vusA);
#ifdef _DEBUG
	for(k=0;k<my_min(TEST_FRAME_LEN,vusA.size());k++){
		vusTmpA[k]=vusA[k];
	}
#endif
	/* remove short segments */
	vus_remove_short_segments(vusA);
	return frameN;
}
int Fe::vus_median_filter(vector<EVusType>& vusA)
{
	int n,i,k;
	int frameN=vusA.size();
	/* two-stage median filtering, length 5 */
	vector<EVusType> vusOrgA=vusA;
	float tmp[5];
	for(n=2;n<frameN-2;n++){	
		for(i=n-2,k=0;i<=n+2;i++,k++) tmp[k]=(float)(vusOrgA[i]);
		vusA[n]=(EVusType)((int)(GetMedian(tmp,5)+0.5));
	}
	vusOrgA=vusA;
	for(n=2;n<frameN-2;n++){	
		for(i=n-2,k=0;i<=n+2;i++,k++) tmp[k]=(float)(vusOrgA[i]);
		vusA[n]=(EVusType)((int)(GetMedian(tmp,5)+0.5));
	}
	return frameN;
}
int Fe::vus_remove_short_segments(vector<EVusType>& vusA)
{
	/* remove short voiced segments */
	vus_remove_short_segments_sub(vusA, FRM_VOICED, 5); // old=7
	/* remove short silent segments */
	vus_remove_short_segments_sub(vusA, FRM_SILENCE, 2); // old=2
	/* remove short unvoiced segments */
	vus_remove_short_segments_sub(vusA, FRM_UNVOICED, 3); // old=4
	return 1;
}
/* remove short segments with duration less than or equal to minDur */
int Fe::vus_remove_short_segments_sub(vector<EVusType>& vusA, EVusType type, int minDur)
{
	int startX=1;
	int n,i,k=0;
	EVusType prevType=vusA[0];
	for(n=1;n<vusA.size();n++){
		if(vusA[n] == type && vusA[n-1] != type){
			startX=n;
		}
		else if(vusA[n] != type && vusA[n-1] == type){
			if(n-startX<=minDur){
				for(i=startX;i<n;i++) vusA[i]=vusA[startX-1];
				k++;
			}
			startX=n;
		}
	}
	if(n-startX<=minDur && vusA[n-1]==type){
		for(i=startX;i<n;i++) vusA[i]=vusA[startX-1];
		k++;
	}
	return k;
}

						