[hts-users:02312] A bug in HHEd.c:FloorAverageCommand()

["Heiga ZEN (Byung Ha CHUN)" <heiga.zen@xxxxxxxxxxxxxxxxx>]

Hi,

I found a bug in HHEd.c:FloorAverageCommand(): it doesn't support multi-stream 
and/or MSD HMMs.  The attached patch code fixes this problem.

Best regards,

Heiga ZEN (Byung Ha CHUN)

--
--------------------------
Heiga ZEN (Byung Ha CHUN)
Speech Technology Group
Cambridge Research Lab
Toshiba Research Europe
phone: +44 1223 436975


______________________________________________________________________
This email has been scanned by the MessageLabs Email Security System.
For more information please visit http://www.messagelabs.com/email 
______________________________________________________________________
--- HHEd.c	2008-08-07 13:36:19.341080000 +0000
+++ HHEd.c	2009-12-14 10:36:38.490135000 +0000
@@ -6572,11 +6681,12 @@
    StreamInfo *sti;
    SVector var , mean;
    DVector *varAcc;
-   double occAcc;
+   double occAcc[SMAX];
    float weight;
-   float occ; 
+   float occ,occs; 
    int l=0,k,i,s,S;
    float varScale;
+   Boolean singleMix;
 
    /*  need stats for operation */
    if (!occStatsLoaded)
@@ -6595,36 +6705,55 @@
       ZeroDVector(varAcc[s]);
    }
    NewHMMScan(hset,&hss);
-   occAcc = 0.0;
+   for (s=1; s<=hset->swidth[0]; s++) occAcc[s] = 0.0;
    while(GoNextState(&hss,FALSE)) {
       si = hss.si;
-      memcpy(&occ,&(si->hook),sizeof(float));
-      occAcc += occ;
-      s=0;
       while (GoNextStream(&hss,TRUE)) {
-         l=hset->swidth[hss.s];
+         s=hss.s;
+         l=hset->swidth[s];
          sti = hss.sti;
-         s++;
-         if ( hss.M > 1 ) { 
+         memcpy(&occ,&(sti->hook),sizeof(float));
+
+         /* single/multiple mix check */
+         for (i=1,k=0; i<=hss.M; i++) {
+            if (hset->swidth[s] == VectorSize(sti->spdf.cpdf[i].mpdf->mean))  /* MSD check */
+               k++;
+         }
+         singleMix = (k==1) ? TRUE : FALSE;
+
+         /* accumulate statistics */
+         if ( !singleMix ) { 
+            occs = 0.0;
             for (k=1;k<=l;k++) { /* loop over k-th dimension */
                double    rvar  = 0.0;
                double    rmean = 0.0;
                for (i=1; i<=hss.M; i++) {
+                  if (VectorSize(sti->spdf.cpdf[i].mpdf->mean) == l) {  /* MSD check */
                   weight = sti->spdf.cpdf[i].weight;
                   var  = sti->spdf.cpdf[i].mpdf->cov.var;
                   mean = sti->spdf.cpdf[i].mpdf->mean;
                   
                   rvar  += weight * ( var[k] + mean[k] * mean[k] );
                   rmean += weight * mean[k];
+
+                     if (k==1) occs += weight * occ;
+                  }
                }
                rvar -= rmean * rmean;
-               varAcc[s][k] += rvar * occ ;
+               varAcc[s][k] += rvar * occs ;
             }
+            occAcc[s] += occs;
          }
-         else { /* single mix */
-            var  = sti->spdf.cpdf[1].mpdf->cov.var;
+         else { /* single mix (single space matches the stream weight) */
+            for (i=1; i<=hss.M; i++) {
+               if (VectorSize(sti->spdf.cpdf[i].mpdf->mean) == l) {
+                  weight = sti->spdf.cpdf[i].weight;
+                  var    = sti->spdf.cpdf[i].mpdf->cov.var;
             for (k=1;k<=l;k++) 
-               varAcc[s][k] += var[k]*occ;
+                     varAcc[s][k] += occ * weight * var[k];
+                  occAcc[s] += occ * weight;
+               }
+            }
          }
       }
    }
@@ -6632,7 +6761,8 @@
    /* normalisation */
    for (s=1;s<=S;s++)
       for (k=1;k<=hset->swidth[s];k++)
-         varAcc[s][k] /= occAcc;
+         varAcc[s][k] /= occAcc[s];
+
    /* set the varFloorN macros */
    for (s=1; s<=S; s++){
       int size;
@@ -6657,7 +6787,7 @@
          NewMacro(hset,hset->numFiles,'v',id,v);
       }
       /* scale and store */
-      for (k=1;k<=l;k++)
+      for (k=1;k<=hset->swidth[s];k++)
          v[k] = varAcc[s][k]*varScale;
    }
    /* and apply floors */