static __inline int32_t Log2Q8( uint32_t x ) {
int32_t zeros, lg2;
int16_t frac;
zeros=WebRtcSpl_NormU32(x);
frac=(int16_t)WEBRTC_SPL_RSHIFT_W32(((uint32_t)(WEBRTC_SPL_LSHIFT_W32(x, zeros))&0x7FFFFFFF), 23);
/* log2(magn(i)) */
lg2= (WEBRTC_SPL_LSHIFT_W32((31-zeros), 8)+frac);
return lg2;
}
static __inline WebRtc_Word32 log2_Q8_LPC( WebRtc_UWord32 x ) {
WebRtc_Word32 zeros, lg2;
WebRtc_Word16 frac;
zeros=WebRtcSpl_NormU32(x);
frac=(WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(((WebRtc_UWord32)WEBRTC_SPL_LSHIFT_W32(x, zeros)&0x7FFFFFFF), 23);
/* log2(x) */
lg2= (WEBRTC_SPL_LSHIFT_W16((31-zeros), 8)+frac);
return lg2;
}
// Noise Estimation
static void NoiseEstimationNeon(NsxInst_t* inst,
uint16_t* magn,
uint32_t* noise,
int16_t* q_noise) {
int16_t lmagn[HALF_ANAL_BLOCKL], counter, countDiv;
int16_t countProd, delta, zeros, frac;
int16_t log2, tabind, logval, tmp16, tmp16no1, tmp16no2;
const int16_t log2_const = 22713;
const int16_t width_factor = 21845;
int i, s, offset;
tabind = inst->stages - inst->normData;
assert(tabind < 9);
assert(tabind > -9);
if (tabind < 0) {
logval = -WebRtcNsx_kLogTable[-tabind];
} else {
logval = WebRtcNsx_kLogTable[tabind];
}
int16x8_t logval_16x8 = vdupq_n_s16(logval);
// lmagn(i)=log(magn(i))=log(2)*log2(magn(i))
// magn is in Q(-stages), and the real lmagn values are:
// real_lmagn(i)=log(magn(i)*2^stages)=log(magn(i))+log(2^stages)
// lmagn in Q8
for (i = 0; i < inst->magnLen; i++) {
if (magn[i]) {
zeros = WebRtcSpl_NormU32((uint32_t)magn[i]);
frac = (int16_t)((((uint32_t)magn[i] << zeros)
& 0x7FFFFFFF) >> 23);
assert(frac < 256);
// log2(magn(i))
log2 = (int16_t)(((31 - zeros) << 8)
+ WebRtcNsx_kLogTableFrac[frac]);
// log2(magn(i))*log(2)
lmagn[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(log2, log2_const, 15);
// + log(2^stages)
lmagn[i] += logval;
} else {
lmagn[i] = logval;
}
}
void WebRtcVad_LogOfEnergy(WebRtc_Word16 *vector,
WebRtc_Word16 *enerlogval,
WebRtc_Word16 *power,
WebRtc_Word16 offset,
int vector_length)
{
WebRtc_Word16 enerSum = 0;
WebRtc_Word16 zeros, frac, log2;
WebRtc_Word32 energy;
int shfts = 0, shfts2;
energy = WebRtcSpl_Energy(vector, vector_length, &shfts);
if (energy > 0)
{
shfts2 = 16 - WebRtcSpl_NormW32(energy);
shfts += shfts2;
// "shfts" is the total number of right shifts that has been done to enerSum.
enerSum = (WebRtc_Word16)WEBRTC_SPL_SHIFT_W32(energy, -shfts2);
// Find:
// 160*log10(enerSum*2^shfts) = 160*log10(2)*log2(enerSum*2^shfts) =
// 160*log10(2)*(log2(enerSum) + log2(2^shfts)) =
// 160*log10(2)*(log2(enerSum) + shfts)
zeros = WebRtcSpl_NormU32(enerSum);
frac = (WebRtc_Word16)(((WebRtc_UWord32)((WebRtc_Word32)(enerSum) << zeros)
& 0x7FFFFFFF) >> 21);
log2 = (WebRtc_Word16)(((31 - zeros) << 10) + frac);
*enerlogval = (WebRtc_Word16)WEBRTC_SPL_MUL_16_16_RSFT(kLogConst, log2, 19)
+ (WebRtc_Word16)WEBRTC_SPL_MUL_16_16_RSFT(shfts, kLogConst, 9);
if (*enerlogval < 0)
{
*enerlogval = 0;
}
} else
// Compute speech/noise probability
// speech/noise probability is returned in: probSpeechFinal
//snrLocPrior is the prior SNR for each frequency (in Q11)
//snrLocPost is the post SNR for each frequency (in Q11)
void WebRtcNsx_SpeechNoiseProb(NsxInst_t* inst,
uint16_t* nonSpeechProbFinal,
uint32_t* priorLocSnr,
uint32_t* postLocSnr) {
uint32_t zeros, num, den, tmpU32no1, tmpU32no2, tmpU32no3;
int32_t invLrtFX, indPriorFX, tmp32, tmp32no1, tmp32no2, besselTmpFX32;
int32_t frac32, logTmp;
int32_t logLrtTimeAvgKsumFX;
int16_t indPriorFX16;
int16_t tmp16, tmp16no1, tmp16no2, tmpIndFX, tableIndex, frac, intPart;
int i, normTmp, normTmp2, nShifts;
// compute feature based on average LR factor
// this is the average over all frequencies of the smooth log LRT
logLrtTimeAvgKsumFX = 0;
for (i = 0; i < inst->magnLen; i++) {
besselTmpFX32 = (int32_t)postLocSnr[i]; // Q11
normTmp = WebRtcSpl_NormU32(postLocSnr[i]);
num = WEBRTC_SPL_LSHIFT_U32(postLocSnr[i], normTmp); // Q(11+normTmp)
if (normTmp > 10) {
den = WEBRTC_SPL_LSHIFT_U32(priorLocSnr[i], normTmp - 11); // Q(normTmp)
} else {
den = WEBRTC_SPL_RSHIFT_U32(priorLocSnr[i], 11 - normTmp); // Q(normTmp)
}
if (den > 0) {
besselTmpFX32 -= WEBRTC_SPL_UDIV(num, den); // Q11
} else {
besselTmpFX32 -= num; // Q11
}
// inst->logLrtTimeAvg[i] += LRT_TAVG * (besselTmp - log(snrLocPrior)
// - inst->logLrtTimeAvg[i]);
// Here, LRT_TAVG = 0.5
zeros = WebRtcSpl_NormU32(priorLocSnr[i]);
frac32 = (int32_t)(((priorLocSnr[i] << zeros) & 0x7FFFFFFF) >> 19);
tmp32 = WEBRTC_SPL_MUL(frac32, frac32);
tmp32 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL(tmp32, -43), 19);
tmp32 += WEBRTC_SPL_MUL_16_16_RSFT((int16_t)frac32, 5412, 12);
frac32 = tmp32 + 37;
// tmp32 = log2(priorLocSnr[i])
tmp32 = (int32_t)(((31 - zeros) << 12) + frac32) - (11 << 12); // Q12
logTmp = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_32_16(tmp32, 178), 8);
// log2(priorLocSnr[i])*log(2)
tmp32no1 = WEBRTC_SPL_RSHIFT_W32(logTmp + inst->logLrtTimeAvgW32[i], 1);
// Q12
inst->logLrtTimeAvgW32[i] += (besselTmpFX32 - tmp32no1); // Q12
logLrtTimeAvgKsumFX += inst->logLrtTimeAvgW32[i]; // Q12
}
inst->featureLogLrt = WEBRTC_SPL_RSHIFT_W32(logLrtTimeAvgKsumFX * 5,
inst->stages + 10);
// 5 = BIN_SIZE_LRT / 2
// done with computation of LR factor
//
//compute the indicator functions
//
// average LRT feature
// FLOAT code
// indicator0 = 0.5 * (tanh(widthPrior *
// (logLrtTimeAvgKsum - threshPrior0)) + 1.0);
tmpIndFX = 16384; // Q14(1.0)
tmp32no1 = logLrtTimeAvgKsumFX - inst->thresholdLogLrt; // Q12
nShifts = 7 - inst->stages; // WIDTH_PR_MAP_SHIFT - inst->stages + 5;
//use larger width in tanh map for pause regions
if (tmp32no1 < 0) {
tmpIndFX = 0;
tmp32no1 = -tmp32no1;
//widthPrior = widthPrior * 2.0;
nShifts++;
}
tmp32no1 = WEBRTC_SPL_SHIFT_W32(tmp32no1, nShifts); // Q14
// compute indicator function: sigmoid map
tableIndex = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32no1, 14);
if ((tableIndex < 16) && (tableIndex >= 0)) {
tmp16no2 = kIndicatorTable[tableIndex];
tmp16no1 = kIndicatorTable[tableIndex + 1] - kIndicatorTable[tableIndex];
frac = (int16_t)(tmp32no1 & 0x00003fff); // Q14
tmp16no2 += (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(tmp16no1, frac, 14);
if (tmpIndFX == 0) {
tmpIndFX = 8192 - tmp16no2; // Q14
} else {
tmpIndFX = 8192 + tmp16no2; // Q14
}
}
indPriorFX = WEBRTC_SPL_MUL_16_16(inst->weightLogLrt, tmpIndFX); // 6*Q14
//spectral flatness feature
if (inst->weightSpecFlat) {
tmpU32no1 = WEBRTC_SPL_UMUL(inst->featureSpecFlat, 400); // Q10
tmpIndFX = 16384; // Q14(1.0)
//use larger width in tanh map for pause regions
tmpU32no2 = inst->thresholdSpecFlat - tmpU32no1; //Q10
nShifts = 4;
if (inst->thresholdSpecFlat < tmpU32no1) {
tmpIndFX = 0;
tmpU32no2 = tmpU32no1 - inst->thresholdSpecFlat;
//widthPrior = widthPrior * 2.0;
nShifts++;
}
tmp32no1 = (int32_t)WebRtcSpl_DivU32U16(WEBRTC_SPL_LSHIFT_U32(tmpU32no2,
nShifts), 25);
//Q14
tmpU32no1 = WebRtcSpl_DivU32U16(WEBRTC_SPL_LSHIFT_U32(tmpU32no2, nShifts),
25); //Q14
// compute indicator function: sigmoid map
// FLOAT code
// indicator1 = 0.5 * (tanh(sgnMap * widthPrior *
// (threshPrior1 - tmpFloat1)) + 1.0);
tableIndex = (int16_t)WEBRTC_SPL_RSHIFT_U32(tmpU32no1, 14);
if (tableIndex < 16) {
tmp16no2 = kIndicatorTable[tableIndex];
tmp16no1 = kIndicatorTable[tableIndex + 1] - kIndicatorTable[tableIndex];
frac = (int16_t)(tmpU32no1 & 0x00003fff); // Q14
tmp16no2 += (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(tmp16no1, frac, 14);
if (tmpIndFX) {
tmpIndFX = 8192 + tmp16no2; // Q14
} else {
tmpIndFX = 8192 - tmp16no2; // Q14
}
}
indPriorFX += WEBRTC_SPL_MUL_16_16(inst->weightSpecFlat, tmpIndFX); // 6*Q14
}
//for template spectral-difference
if (inst->weightSpecDiff) {
tmpU32no1 = 0;
if (inst->featureSpecDiff) {
normTmp = WEBRTC_SPL_MIN(20 - inst->stages,
WebRtcSpl_NormU32(inst->featureSpecDiff));
tmpU32no1 = WEBRTC_SPL_LSHIFT_U32(inst->featureSpecDiff, normTmp);
// Q(normTmp-2*stages)
tmpU32no2 = WEBRTC_SPL_RSHIFT_U32(inst->timeAvgMagnEnergy,
20 - inst->stages - normTmp);
if (tmpU32no2 > 0) {
// Q(20 - inst->stages)
tmpU32no1 = WEBRTC_SPL_UDIV(tmpU32no1, tmpU32no2);
} else {
tmpU32no1 = (uint32_t)(0x7fffffff);
}
}
tmpU32no3 = WEBRTC_SPL_UDIV(WEBRTC_SPL_LSHIFT_U32(inst->thresholdSpecDiff,
17),
25);
tmpU32no2 = tmpU32no1 - tmpU32no3;
nShifts = 1;
tmpIndFX = 16384; // Q14(1.0)
//use larger width in tanh map for pause regions
if (tmpU32no2 & 0x80000000) {
tmpIndFX = 0;
tmpU32no2 = tmpU32no3 - tmpU32no1;
//widthPrior = widthPrior * 2.0;
nShifts--;
}
tmpU32no1 = WEBRTC_SPL_RSHIFT_U32(tmpU32no2, nShifts);
// compute indicator function: sigmoid map
/* FLOAT code
indicator2 = 0.5 * (tanh(widthPrior * (tmpFloat1 - threshPrior2)) + 1.0);
*/
tableIndex = (int16_t)WEBRTC_SPL_RSHIFT_U32(tmpU32no1, 14);
if (tableIndex < 16) {
tmp16no2 = kIndicatorTable[tableIndex];
tmp16no1 = kIndicatorTable[tableIndex + 1] - kIndicatorTable[tableIndex];
frac = (int16_t)(tmpU32no1 & 0x00003fff); // Q14
tmp16no2 += (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(
tmp16no1, frac, 14);
if (tmpIndFX) {
tmpIndFX = 8192 + tmp16no2;
} else {
tmpIndFX = 8192 - tmp16no2;
}
}
indPriorFX += WEBRTC_SPL_MUL_16_16(inst->weightSpecDiff, tmpIndFX); // 6*Q14
}
//combine the indicator function with the feature weights
// FLOAT code
// indPrior = 1 - (weightIndPrior0 * indicator0 + weightIndPrior1 *
// indicator1 + weightIndPrior2 * indicator2);
indPriorFX16 = WebRtcSpl_DivW32W16ResW16(98307 - indPriorFX, 6); // Q14
// done with computing indicator function
//compute the prior probability
// FLOAT code
// inst->priorNonSpeechProb += PRIOR_UPDATE *
// (indPriorNonSpeech - inst->priorNonSpeechProb);
tmp16 = indPriorFX16 - inst->priorNonSpeechProb; // Q14
inst->priorNonSpeechProb += (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(
PRIOR_UPDATE_Q14, tmp16, 14); // Q14
//final speech probability: combine prior model with LR factor:
memset(nonSpeechProbFinal, 0, sizeof(uint16_t) * inst->magnLen);
if (inst->priorNonSpeechProb > 0) {
for (i = 0; i < inst->magnLen; i++) {
// FLOAT code
// invLrt = exp(inst->logLrtTimeAvg[i]);
// invLrt = inst->priorSpeechProb * invLrt;
// nonSpeechProbFinal[i] = (1.0 - inst->priorSpeechProb) /
// (1.0 - inst->priorSpeechProb + invLrt);
// invLrt = (1.0 - inst->priorNonSpeechProb) * invLrt;
// nonSpeechProbFinal[i] = inst->priorNonSpeechProb /
// (inst->priorNonSpeechProb + invLrt);
if (inst->logLrtTimeAvgW32[i] < 65300) {
tmp32no1 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL(
inst->logLrtTimeAvgW32[i], 23637),
14); // Q12
intPart = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32no1, 12);
if (intPart < -8) {
intPart = -8;
}
frac = (int16_t)(tmp32no1 & 0x00000fff); // Q12
// Quadratic approximation of 2^frac
tmp32no2 = WEBRTC_SPL_RSHIFT_W32(frac * frac * 44, 19); // Q12
tmp32no2 += WEBRTC_SPL_MUL_16_16_RSFT(frac, 84, 7); // Q12
invLrtFX = WEBRTC_SPL_LSHIFT_W32(1, 8 + intPart)
+ WEBRTC_SPL_SHIFT_W32(tmp32no2, intPart - 4); // Q8
normTmp = WebRtcSpl_NormW32(invLrtFX);
normTmp2 = WebRtcSpl_NormW16((16384 - inst->priorNonSpeechProb));
if (normTmp + normTmp2 >= 7) {
if (normTmp + normTmp2 < 15) {
invLrtFX = WEBRTC_SPL_RSHIFT_W32(invLrtFX, 15 - normTmp2 - normTmp);
// Q(normTmp+normTmp2-7)
tmp32no1 = WEBRTC_SPL_MUL_32_16(invLrtFX,
(16384 - inst->priorNonSpeechProb));
// Q(normTmp+normTmp2+7)
invLrtFX = WEBRTC_SPL_SHIFT_W32(tmp32no1, 7 - normTmp - normTmp2);
// Q14
} else {
tmp32no1 = WEBRTC_SPL_MUL_32_16(invLrtFX,
(16384 - inst->priorNonSpeechProb));
// Q22
invLrtFX = WEBRTC_SPL_RSHIFT_W32(tmp32no1, 8); // Q14
}
tmp32no1 = WEBRTC_SPL_LSHIFT_W32((int32_t)inst->priorNonSpeechProb,
8); // Q22
nonSpeechProbFinal[i] = (uint16_t)WEBRTC_SPL_DIV(tmp32no1,
(int32_t)inst->priorNonSpeechProb + invLrtFX); // Q8
}
}
}
}
}
