/* 1D parabolic interpolation . All input and output values are in Q8 */
static __inline void Intrp1DQ8(int32_t *x, int32_t *fx, int32_t *y, int32_t *fy) {
int16_t sign1=1, sign2=1;
int32_t r32, q32, t32, nom32, den32;
int16_t t16, tmp16, tmp16_1;
if ((fx[0]>0) && (fx[2]>0)) {
r32=fx[1]-fx[2];
q32=fx[0]-fx[1];
nom32=q32+r32;
den32=WEBRTC_SPL_MUL_32_16((q32-r32), 2);
if (nom32<0)
sign1=-1;
if (den32<0)
sign2=-1;
/* t = (q32+r32)/(2*(q32-r32)) = (fx[0]-fx[1] + fx[1]-fx[2])/(2 * fx[0]-fx[1] - (fx[1]-fx[2]))*/
/* (Signs are removed because WebRtcSpl_DivResultInQ31 can't handle negative numbers) */
t32=WebRtcSpl_DivResultInQ31(WEBRTC_SPL_MUL_32_16(nom32, sign1),WEBRTC_SPL_MUL_32_16(den32, sign2)); /* t in Q31, without signs */
t16=(int16_t)WEBRTC_SPL_RSHIFT_W32(t32, 23); /* Q8 */
t16=t16*sign1*sign2; /* t in Q8 with signs */
*y = x[0]+t16; /* Q8 */
// *y = x[1]+t16; /* Q8 */
/* The following code calculates fy in three steps */
/* fy = 0.5 * t * (t-1) * fx[0] + (1-t*t) * fx[1] + 0.5 * t * (t+1) * fx[2]; */
/* Part I: 0.5 * t * (t-1) * fx[0] */
tmp16_1=(int16_t)WEBRTC_SPL_MUL_16_16(t16,t16); /* Q8*Q8=Q16 */
tmp16_1 = WEBRTC_SPL_RSHIFT_W16(tmp16_1,2); /* Q16>>2 = Q14 */
t16 = (int16_t)WEBRTC_SPL_MUL_16_16(t16, 64); /* Q8<<6 = Q14 */
tmp16 = tmp16_1-t16;
*fy = WEBRTC_SPL_MUL_16_32_RSFT15(tmp16, fx[0]); /* (Q14 * Q8 >>15)/2 = Q8 */
/* Part II: (1-t*t) * fx[1] */
tmp16 = 16384-tmp16_1; /* 1 in Q14 - Q14 */
*fy += WEBRTC_SPL_MUL_16_32_RSFT14(tmp16, fx[1]);/* Q14 * Q8 >> 14 = Q8 */
/* Part III: 0.5 * t * (t+1) * fx[2] */
tmp16 = tmp16_1+t16;
*fy += WEBRTC_SPL_MUL_16_32_RSFT15(tmp16, fx[2]);/* (Q14 * Q8 >>15)/2 = Q8 */
} else {
*y = x[0];
*fy= fx[1];
}
}
WebRtc_UWord32 WebRtcNetEQ_ScaleTimestampInternalToExternal(const MCUInst_t *MCU_inst,
WebRtc_UWord32 internalTS)
{
WebRtc_Word32 timestampDiff;
WebRtc_UWord32 externalTS;
/* difference between this and last incoming timestamp */
timestampDiff = (WebRtc_Word32) internalTS - MCU_inst->internalTS;
switch (MCU_inst->scalingFactor)
{
case kTSscalingTwo:
{
/* divide by 2 */
timestampDiff = WEBRTC_SPL_RSHIFT_W32(timestampDiff, 1);
break;
}
case kTSscalingTwoThirds:
{
/* multiply with 3/2 */
timestampDiff = WEBRTC_SPL_MUL_32_16(timestampDiff, 3);
timestampDiff = WEBRTC_SPL_RSHIFT_W32(timestampDiff, 1);
break;
}
case kTSscalingFourThirds:
{
/* multiply with 3/4 */
timestampDiff = WEBRTC_SPL_MUL_32_16(timestampDiff, 3);
timestampDiff = WEBRTC_SPL_RSHIFT_W32(timestampDiff, 2);
break;
}
default:
{
/* no scaling */
}
}
/* add the scaled difference to last scaled timestamp and save ... */
externalTS = MCU_inst->externalTS + timestampDiff;
return externalTS;
}
int WebRtcNetEQ_UpdateIatStatistics(AutomodeInst_t *inst, int maxBufLen,
uint16_t seqNumber, uint32_t timeStamp,
int32_t fsHz, int mdCodec, int streamingMode)
{
uint32_t timeIat; /* inter-arrival time */
int i;
int32_t tempsum = 0; /* temp summation */
int32_t tempvar; /* temporary variable */
int retval = 0; /* return value */
int16_t packetLenSamp; /* packet speech length in samples */
/****************/
/* Sanity check */
/****************/
if (maxBufLen <= 1 || fsHz <= 0)
{
/* maxBufLen must be at least 2 and fsHz must both be strictly positive */
return -1;
}
/****************************/
/* Update packet statistics */
/****************************/
/* Try calculating packet length from current and previous timestamps */
if (!WebRtcNetEQ_IsNewerTimestamp(timeStamp, inst->lastTimeStamp) ||
!WebRtcNetEQ_IsNewerSequenceNumber(seqNumber, inst->lastSeqNo))
{
/* Wrong timestamp or sequence order; revert to backup plan */
packetLenSamp = inst->packetSpeechLenSamp; /* use stored value */
}
else
{
/* calculate timestamps per packet */
packetLenSamp = (int16_t) WebRtcSpl_DivU32U16(timeStamp - inst->lastTimeStamp,
seqNumber - inst->lastSeqNo);
}
/* Check that the packet size is positive; if not, the statistics cannot be updated. */
if (inst->firstPacketReceived && packetLenSamp > 0)
{ /* packet size ok */
/* calculate inter-arrival time in integer packets (rounding down) */
timeIat = WebRtcSpl_DivW32W16(inst->packetIatCountSamp, packetLenSamp);
/* Special operations for streaming mode */
if (streamingMode != 0)
{
/*
* Calculate IAT in Q8, including fractions of a packet (i.e., more accurate
* than timeIat).
*/
int16_t timeIatQ8 = (int16_t) WebRtcSpl_DivW32W16(
WEBRTC_SPL_LSHIFT_W32(inst->packetIatCountSamp, 8), packetLenSamp);
/*
* Calculate cumulative sum iat with sequence number compensation (ideal arrival
* times makes this sum zero).
*/
inst->cSumIatQ8 += (timeIatQ8
- WEBRTC_SPL_LSHIFT_W32(seqNumber - inst->lastSeqNo, 8));
/* subtract drift term */
inst->cSumIatQ8 -= CSUM_IAT_DRIFT;
/* ensure not negative */
inst->cSumIatQ8 = WEBRTC_SPL_MAX(inst->cSumIatQ8, 0);
/* remember max */
if (inst->cSumIatQ8 > inst->maxCSumIatQ8)
{
inst->maxCSumIatQ8 = inst->cSumIatQ8;
inst->maxCSumUpdateTimer = 0;
}
/* too long since the last maximum was observed; decrease max value */
if (inst->maxCSumUpdateTimer > (uint32_t) WEBRTC_SPL_MUL_32_16(fsHz,
MAX_STREAMING_PEAK_PERIOD))
{
inst->maxCSumIatQ8 -= 4; /* remove 1000*4/256 = 15.6 ms/s */
}
} /* end of streaming mode */
/* check for discontinuous packet sequence and re-ordering */
if (WebRtcNetEQ_IsNewerSequenceNumber(seqNumber, inst->lastSeqNo + 1))
{
/* Compensate for gap in the sequence numbers.
* Reduce IAT with expected extra time due to lost packets, but ensure that
* the IAT is not negative.
*/
timeIat -= WEBRTC_SPL_MIN(timeIat,
(uint16_t) (seqNumber - (uint16_t) (inst->lastSeqNo + 1)));
}
else if (!WebRtcNetEQ_IsNewerSequenceNumber(seqNumber, inst->lastSeqNo))
{
/* compensate for re-ordering */
timeIat += (uint16_t) (inst->lastSeqNo + 1 - seqNumber);
}
/* saturate IAT at maximum value */
timeIat = WEBRTC_SPL_MIN( timeIat, MAX_IAT );
/* update iatProb = forgetting_factor * iatProb for all elements */
for (i = 0; i <= MAX_IAT; i++)
{
int32_t tempHi, tempLo; /* Temporary variables */
/*
* Multiply iatProbFact (Q15) with iatProb (Q30) and right-shift 15 steps
* to come back to Q30. The operation is done in two steps:
*/
/*
* 1) Multiply the high 16 bits (15 bits + sign) of iatProb. Shift iatProb
* 16 steps right to get the high 16 bits in a int16_t prior to
* multiplication, and left-shift with 1 afterwards to come back to
* Q30 = (Q15 * (Q30>>16)) << 1.
*/
tempHi = WEBRTC_SPL_MUL_16_16(inst->iatProbFact,
(int16_t) WEBRTC_SPL_RSHIFT_W32(inst->iatProb[i], 16));
tempHi = WEBRTC_SPL_LSHIFT_W32(tempHi, 1); /* left-shift 1 step */
/*
* 2) Isolate and multiply the low 16 bits of iatProb. Right-shift 15 steps
* afterwards to come back to Q30 = (Q15 * Q30) >> 15.
*/
tempLo = inst->iatProb[i] & 0x0000FFFF; /* sift out the 16 low bits */
tempLo = WEBRTC_SPL_MUL_16_U16(inst->iatProbFact,
(uint16_t) tempLo);
tempLo = WEBRTC_SPL_RSHIFT_W32(tempLo, 15);
/* Finally, add the high and low parts */
inst->iatProb[i] = tempHi + tempLo;
/* Sum all vector elements while we are at it... */
tempsum += inst->iatProb[i];
}
/*
* Increase the probability for the currently observed inter-arrival time
* with 1 - iatProbFact. The factor is in Q15, iatProb in Q30;
* hence, left-shift 15 steps to obtain result in Q30.
*/
inst->iatProb[timeIat] += (32768 - inst->iatProbFact) << 15;
tempsum += (32768 - inst->iatProbFact) << 15; /* add to vector sum */
/*
* Update iatProbFact (changes only during the first seconds after reset)
* The factor converges to IAT_PROB_FACT.
*/
inst->iatProbFact += (IAT_PROB_FACT - inst->iatProbFact + 3) >> 2;
/* iatProb should sum up to 1 (in Q30). */
tempsum -= 1 << 30; /* should be zero */
/* Check if it does, correct if it doesn't. */
if (tempsum > 0)
{
/* tempsum too large => decrease a few values in the beginning */
i = 0;
while (i <= MAX_IAT && tempsum > 0)
{
/* Remove iatProb[i] / 16 from iatProb, but not more than tempsum */
tempvar = WEBRTC_SPL_MIN(tempsum, inst->iatProb[i] >> 4);
inst->iatProb[i++] -= tempvar;
tempsum -= tempvar;
}
}
// 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
}
}
}
}
}
