语音识别

Hmm算法，nuance asr关键实现 


hmm为目前最流行的大规模语音识别算法，基本思想是把每个声音片段（20ms）的特征信息（如MFCC）作为每个马科夫链的观察节点，说话过程，就是这些节点之间的转移，根据已经了解到每个音节对应的节点跳转关系，可以识别出当前语言对应的最靠谱的音节（ 维特比算法）。







//------------------------------------------------------


// calculate the log prob at a state with given ob


// It doesnt matter if alpha not used in denominator!


//------------------------------------------------------


float CalculateLogProb(STATE *s, float *ob, int veclen)


{


int m, n, nVecLen, nMix;


float alpha, det, tmp, sum;


float prob=0.0;






if(!s || !ob) return prob;


// if we can use pre-computed constant?


nVecLen = veclen;


if(nVecLen != VecLength) {


alpha = (float)pow(TWOPI, nVecLen/2.0);


AlphaConst = alpha;


}


else {


alpha = AlphaConst;


}


// for each mixture, compute prob


nMix = s->nMixtures;


for(m=0; m<nMix; m++) {


sum = 0.0;


det = 1.0;


for(n=0; n<nVecLen; n++){


tmp  = ob[n] - (s->pfMean[m][n]);


sum += tmp*tmp/(s->pfCov[m][n]);


det *= (float)(s->pfCov[m][n]);


}


det = (float)sqrt(det);


prob += (s->pfMixWt[m])*exp(-0.5*sum)/(alpha*det);


}


// check if prob is valid?


if( prob < (float)PROBFLOOR )


prob = (float)PROBFLOOR;


prob = (float)log(prob);






return prob;


}