📄 encoder.c
字号:
ippsDotProd_16s32s_Sfs(pBuffer,pBuffer,FRAMESHIFT,&sum,0);
sum+=64;
ippsLn_32s_ISfs(&sum, 1, -10);
frameEn = (Ipp32s)(512 + ((23637 * (Ipp64s)(sum-4259))>>10));
if( ((frameEn-vad->Eloglowtrack) < SNR_THRESHOLD_UPD_LTE) || (tc < MIN_FRAME) ){
if( (frameEn<vad->Eloglowtrack) || ((tc) < MIN_FRAME))
vad->Eloglowtrack = vad->Eloglowtrack + ((Ipp32s)lambdaLTE *( frameEn - vad->Eloglowtrack )>>10);
else
vad->Eloglowtrack = vad->Eloglowtrack + (10*( frameEn - vad->Eloglowtrack )>>10);
if (vad->Eloglowtrack < ENERGY_FLOOR)vad->Eloglowtrack = ENERGY_FLOOR;
}
if(tc>4){
if( ( (frameEn-vad->Eloglowtrack) > SNR_THRESHOLD_VAD) ){
vad->nbSpeechFrame++;
vad->flagVAD = 1;
}
else {
if( vad->nbSpeechFrame > MIN_SPEECH_FRAME_HANGOVER)
vad->hangOver = HANGOVER;
vad->nbSpeechFrame = 0;
if(vad->hangOver!=0){
vad->hangOver--;
vad->flagVAD = 1;
}
else
vad->flagVAD = 0;
}
}
return vad->flagVAD;
}
void DoNoiseReduction(AuroraEncoder *pCodec,short *pBuffer, int tc, int *pOutFFT, short *pOutputBuffer, int *tcur,
short *pOutMel, short *pOutDen,char flagVAD){
int i;
Ipp32s pFFT32s[FFTSIZE+2];
Ipp16s pFFT16s[FFTSIZE+2];
Ipp64s pFFT64s[FFTSIZE+2];
Ipp64s pPSD64s[FFTSPEC];
Ipp16s pPSD16s[FFTSPEC];
Ipp32s pPSD32s[FFTSPEC];
Ipp16s pNoise16s[FFTSPEC];
Ipp32s pNoise32s[FFTSPEC];
Ipp16s pDen16s[FFTSPEC];
Ipp16s pDataOut16s[200];
/* Gain factorization */
Ipp32s SNRlow32s = 0;
Ipp32s agf=0;
Ipp32s Enoise32s;
Ipp32s SNRaver32s;
Ipp32f lambdaNSE;
Ipp32f lsnr;
/* Mel filter-bank */
Ipp16s pMelBank16s[NUM_CHANNELS+2];
Ipp16s phWF16s[(NUM_CHANNELS+1)*2+1];
Ipp16s phWFtrunc16s[FL+1];
Ipp16s *pState2 = pBuffer+320;
Ipp16s *pState1 = pState2+320;
/*!!!! FIRST STAGE !!!!*/
/* Spectrum estimation */
ippsMul_16s32s_Sfs(winHannCoeffQ14, pState1+60,pFFT32s, WINHANNSIZE,13); /* +4 */
ippsZero_16s((Ipp16s*)(pFFT32s+WINHANNSIZE), (FFTSIZE-WINHANNSIZE)<<1);
ippsFFTFwd_RToCCS_32s_Sfs(pFFT32s,pFFT32s,pCodec->noise.ctxFFT,0,NULL);
pFFT32s[1] = pFFT32s[FFTSIZE];
ippsSmoothedPowerSpectrumAurora_32s_Sfs(pFFT32s,pFFT32s,FFTSIZE,8); /* 0 */
for(i=0;i<FFTSPEC;i++){
pOutFFT[i] = pFFT32s[i];
}
/* Power spectral density mean */
for(i=0;i<FFTSPEC;i++){
pFFT64s[i] = pFFT32s[i];
pPSD64s[i] = (pFFT32s[i] + pCodec->noise.pPSDprev_1[i])>>1;
}
ippsCopy_16s((Ipp16s*)pFFT32s,(Ipp16s*)pCodec->noise.pPSDprev_1,FFTSPEC<<1);
ippsSqrt_64s_ISfs(pPSD64s,FFTSPEC,-2); /* +2 */
ippsSqrt_64s_ISfs(pFFT64s,FFTSPEC,0); /* 0 */
for(i=0;i<FFTSPEC;i++){
pPSD16s[i] = SATURATE_16s(pPSD64s[i]);
pFFT16s[i] = SATURATE_16s(pFFT64s[i]);
}
/* Wiener filter design */
if( tc+1 < NB_FRAME_THRESHOLD_NSE)
lambdaNSE = pLambdaNSE[tc-1];
else
lambdaNSE = LAMBDA_NSE;
if(!flagVAD){
ippsWeightedSum_16s(pCodec->noise.pNoiseTnPrev_1,pPSD16s,pCodec->noise.pNoiseTn,FFTSPEC,lambdaNSE,(1-lambdaNSE)); /* +2 */
ippsThreshold_16s_I(pCodec->noise.pNoiseTn,FFTSPEC,0,ippCmpLess);
}
ippsCopy_16s(pCodec->noise.pNoiseTn,pNoise16s,FFTSPEC);
ippsCopy_16s(pCodec->noise.pNoiseTn,pCodec->noise.pNoiseTnPrev_1,FFTSPEC);
/*pPSD Q4*/
ippsWienerFilterDesign_Aurora_16s(pPSD16s,pNoise16s,pCodec->noise.pDenPrev_1,pDen16s,FFTSPEC); /* +14 */
ippsMul_16s_Sfs(pFFT16s, pDen16s, pCodec->noise.pDenPrev_1, FFTSPEC,12); /* +2 */
pCodec->noise.Eden2 = pCodec->noise.Eden1;pCodec->noise.Eden1 = pCodec->noise.Eden;
ippsSum_16s32s_Sfs(pCodec->noise.pDenPrev_1,FFTSPEC,&pCodec->noise.Eden,2); /* 0 */
/* Mel filter-bank */
ippsEvalFBank_16s_Sfs(pDen16s,pMelBank16s,pCodec->noise.pFBank,-4); /* +18 */
ippsCopy_16s(pDen16s,pOutDen,FFTSPEC);
ippsRShiftC_16s(pMelBank16s,4,pOutMel,NUM_CHANNELS+2);
/* Mel IDCT */
ippsMatVecMul_16s_D2LSfs((const Ipp16s**)pCodec->noise.pIDCT,pMelBank16s,NUM_CHANNELS+2,phWF16s,10,19); /* +14*/
/* Apply filter and forming second buffer */
ippsTabsCalculation_Aurora_16s(phWF16s,phWFtrunc16s);
ippsMul_16s_ISfs(winHannFilter16s, phWFtrunc16s, FL, 10);
ippsResidualFilter_Aurora_16s_Sfs(pState1+FRAMESHIFT-8,pState2+FRAMESHIFT*3,
phWFtrunc16s,13); /*+1*/ /*13*/
/*!!! Start 2nd state !!!*/
if(tc>2){
/* Spectrum estimation 2 */
ippsMul_16s32s_Sfs(winHannCoeffQ14, pState2+60,pFFT32s, WINHANNSIZE,8); /* +6 */
ippsZero_16s((Ipp16s*)(pFFT32s+WINHANNSIZE), (FFTSIZE-WINHANNSIZE)<<1);
ippsFFTFwd_RToCCS_32s_Sfs(pFFT32s,pFFT32s,pCodec->noise.ctxFFT,0,NULL);
pFFT32s[1] = pFFT32s[FFTSIZE];
ippsSmoothedPowerSpectrumAurora_32s_Sfs(pFFT32s,pFFT32s,FFTSIZE,4); /* +8 */
/* Power spectral density mean 2 */
for(i=0;i<FFTSPEC;i++){
pPSD32s[i] = SATURATE_32s((pFFT32s[i] + pCodec->noise.pPSDprev_2[i])>>1); /* +8 */
}
ippsCopy_16s((Ipp16s*)pFFT32s,(Ipp16s*)pCodec->noise.pPSDprev_2,FFTSPEC<<1);
for(i=0;i<FFTSPEC;i++){
pFFT64s[i] = pFFT32s[i];
}
ippsSqrt_64s_ISfs(pFFT64s,FFTSPEC,0); /* +4 */
for(i=0;i<FFTSPEC;i++){
pFFT16s[i] = SATURATE_16s(pFFT64s[i]);
}
/* Wiener filter design 2*/
if( tc+1 < NB_FRAME_THRESHOLD_NSE)
lambdaNSE = pLambdaNSE[tc-3];
else
lambdaNSE = LAMBDA_NSE;
if( tc-2 < 11) {
lambdaNSE = pLambdaNSE[tc-3];
for(i=0;i<FFTSPEC;i++){
pNoise32s[i] = (Ipp32s)(pCodec->noise.pNoisePrev_2[i] * lambdaNSE + pPSD32s[i]*(1-lambdaNSE));
}
/* ippsWeightedSum_16s(pCodec->noise.pNoisePrev_2,pPSD,pNoise,FFTSPEC,lambdaNSE,(1-lambdaNSE)); */
}else{
ippsNoiseSpectrumUpdate_Aurora_32s_Sfs(pPSD32s,pCodec->noise.pNoisePrev_2,pNoise32s,FFTSPEC,0);
}
ippsCopy_16s((Ipp16s*)pNoise32s,(Ipp16s*)pCodec->noise.pNoisePrev_2,FFTSPEC<<1);
for(i=0;i<FFTSPEC;i++){pPSD64s[i] = pPSD32s[i];}
ippsSqrt_64s_ISfs(pPSD64s,FFTSPEC,0);
for(i=0;i<FFTSPEC;i++){pPSD16s[i] = SATURATE_16s(pPSD64s[i]);} /* +4 */
for(i=0;i<FFTSPEC;i++){pPSD64s[i] = pNoise32s[i];}
ippsSqrt_64s_ISfs(pPSD64s,FFTSPEC,0);
for(i=0;i<FFTSPEC;i++){pNoise16s[i] = SATURATE_16s(pPSD64s[i]);}
ippsWienerFilterDesign_Aurora_16s(pPSD16s,pNoise16s,pCodec->noise.pDenPrev_2,pDen16s,FFTSPEC);
ippsMul_16s_Sfs(pFFT16s, pDen16s, pCodec->noise.pDenPrev_2, FFTSPEC,14); /* +4 */
/* Mel filter-bank 2 */
ippsEvalFBank_16s_Sfs(pDen16s,pMelBank16s,pCodec->noise.pFBank,0); /* +14 */
/* Gain factorization */
ippsSum_16s32s_Sfs(pNoise16s,FFTSPEC,&Enoise32s,4); /* 0 */
SNRaver32s=0;
if(Enoise32s!=0)
SNRaver32s = (Ipp32s)(( ((Ipp64s)pCodec->noise.Eden2 * (Ipp64s)pCodec->noise.Eden1 * (Ipp64s)pCodec->noise.Eden)<<21)/
((Ipp64s)Enoise32s * (Ipp64s)Enoise32s * (Ipp64s)Enoise32s));
else{
SNRaver32s = 0;
}
if (SNRaver32s > 21){
ipps10Log10_32s_ISfs(&SNRaver32s,1,-11);
SNRaver32s = 2 * (SNRaver32s - 129467);
}
else{
SNRaver32s = 204800;
}
if( ((SNRaver32s-pCodec->noise.SNRlowPrev32s)<61440) || (tc-2) < 10){
if((tc-2)<10)
lsnr = 1.0f - 1.0f/(tc-2);
else{
if(SNRaver32s < SNRlow32s)
lsnr = 0.95f;
else
lsnr = 0.99f;
}
SNRlow32s = (Ipp32s)(lsnr * pCodec->noise.SNRlowPrev32s + (1-lsnr) * SNRaver32s);
pCodec->noise.SNRlowPrev32s = SNRlow32s;
}
if(SNRaver32s < (SNRlow32s+21504)){
agf = pCodec->noise.agfPrev32s+154;
if(agf > 819)
agf = 819;
}else{
agf = pCodec->noise.agfPrev32s-307;
if(agf < 102)agf = 102;
}
pCodec->noise.agfPrev32s = agf;
for (i=0; i <=NUM_CHANNELS+1; i++){
pMelBank16s[i] = (Ipp16s)((1<<14) + ((agf * (Ipp32s)(pMelBank16s[i]-(1<<14)))>>10));
}
/* Mel IDCT 2 */
ippsMatVecMul_16s_D2LSfs((const Ipp16s**)pCodec->noise.pIDCT,pMelBank16s,
NUM_CHANNELS+2,phWF16s,10,14); /* +15 */
/* Apply filter 2 */
ippsTabsCalculation_Aurora_16s(phWF16s,phWFtrunc16s);
ippsMul_16s_ISfs(winHannFilter16s, phWFtrunc16s, FL,10); /* +15 */
ippsResidualFilter_Aurora_16s_Sfs(pState2+FRAMESHIFT-8,pBuffer+FRAMESHIFT*2,phWFtrunc16s,14); /*+7*/
/* Offset compensation */
ippsCompensateOffset_16s_I(pBuffer+FRAMESHIFT*2,FRAMESHIFT,
&(pCodec->noise.pSrc0),pCodec->noise.Dst0,0.9990234375f);
pCodec->noise.Dst0 = pBuffer[3*FRAMESHIFT-1];
if(tc > 4){
ippsWaveProcessing_Aurora_16s(pBuffer,pDataOut16s);
/* output result */
for(i=0; i<200;i++)pDataOut16s[i] = pDataOut16s[i]>>1;
ippsCopy_16s(pDataOut16s,pOutputBuffer,200);
(*tcur)++;
}
}
}
void ResetAuroraEncoder(AuroraEncoder *pCodec){
pCodec->Dst0=0;
pCodec->preFloat=0;
pCodec->pSrc0=0;
pCodec->curPosition = 0;
pCodec->iCountFrame = 0;
pCodec->mframeCounter = 1;
pCodec->pFrame->pos=6;
pCodec->NumberFrame = 0;
pCodec->VadFrame = 0;
ippsZero_16s(pCodec->bias,NUM_CEP_COEFF-1);
ippsZero_8u(pCodec->pIndexVQBuffer,2*NUM_CODEBOOK);
/* For noice reduction block */
ippsZero_16s((Ipp16s*)pCodec->noise.pPSDprev_1,FFTSPEC<<1);
ippsZero_16s((Ipp16s*)pCodec->noise.pPSDprev_2,FFTSPEC<<1);
ippsZero_16s((Ipp16s*)pCodec->noise.pDenPrev_1,FFTSPEC);
ippsZero_16s((Ipp16s*)pCodec->noise.pDenPrev_2,FFTSPEC);
ippsSet_16s(0,pCodec->noise.pNoiseTnPrev_1,FFTSPEC);
ippsSet_16s(0,(Ipp16s*)pCodec->noise.pNoisePrev_2,FFTSPEC<<1);
ippsSet_16s(0,pCodec->noise.pNoiseTn,FFTSPEC);
ippsZero_16s(pCodec->noise.pBStateBuf, 960);
ippsZero_16s(pCodec->noise.pBStateBig,640+FILTERLENGTH);
pCodec->noise.Eden=pCodec->noise.Eden1=pCodec->noise.Eden2=0;
pCodec->noise.SNRlowPrev32s=0;
pCodec->noise.agfPrev32s=819;
pCodec->noise.pSrc0 = pCodec->noise.Dst0=0;
pCodec->noise.Cur_TPrpocess=0;
pCodec->noise.tail=0;
pCodec->noise.start=0;
pCodec->noise.end=0;
pCodec->noise.vad.nbSpeechFrame=0;
pCodec->noise.vad.Eloglowtrack=0;
pCodec->noise.vad.flagVAD=0;
pCodec->noise.vad.hangOver = 0;
ippsZero_16s(pCodec->noise.s16.dataHP,560);
/* for 16Khz processing */
pCodec->noise.s16.vad16.nbSpeechFrame=0;
pCodec->noise.s16.vad16.Eloglowtrack=0;
pCodec->noise.s16.vad16.flagVAD=0;
pCodec->noise.s16.vad16.hangOver = 0;
pCodec->noise.s16.vad16.pNoise16[0]=0;
pCodec->noise.s16.vad16.pNoise16[1]=0;
pCodec->noise.s16.vad16.pNoise16[2]=0;
pCodec->s11.time = (Ipp64f)pCodec->s11.history;
pCodec->s11.lastread = pCodec->s11.history;
ippsZero_16s((short*)pCodec->s11.p11,+pCodec->s11.len);
}
int InitAuroraEncoder(AuroraEncoder **pCodec, AuroraRate SamplingFrequency,AuroraDataType EncoderInput,AuroraDataType EncoderOutput){
IppStatus status;
int FrameShift=0;
int FrameLength=0;
float startingFeq=0,Frequency;
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -