hvite.c

Hidden Markov Toolkit (HTK) 3.2.1 HTK is a toolkit for use in research into automatic speech recogn

      if (hset.hsKind==DISCRETEHS){
         for (s=1; s<=hset.swidth[0]; s++){
            if( (obs.vq[s] < 1) || (obs.vq[s] > maxMixInS[s]))
               HError(3250,"ProcessFile: Discrete data value [ %d ] out of range in stream [ %d ] in file %s",obs.vq[s],s,fn);
         }
      }

      ProcessObservation(vri,&obs,-1,xfInfo.inXForm);
      
      if (trace & T_FRS) {
         for (d=vri->genMaxNode,j=0;j<30;d=d->links[0].node,j++)
            if (d->type==n_word) break;
         if (d->type==n_word){
            if (d->info.pron==NULL) p=":bound:";
            else p=d->info.pron->word->wordName->name;
         }
         else p=":external:";
         m=FindMacroStruct(&hset,'h',vri->genMaxNode->info.hmm);
         printf("Optimum @%-4d HMM: %s (%s)  %d %5.3f\n",
                vri->frame,m->id->name,p,
                vri->nact,vri->genMaxTok.like/vri->frame);
         fflush(stdout);
      }
      nFrames++;
      tact+=vri->nact;
   }
   lat=CompleteRecognition(vri,pbinfo.tgtSampRate/10000000.0,&ansHeap);
   
   if (lat==NULL) {
      if ((trace & T_TOP) && fn != NULL){
         if (restartable)
            printf("No tokens survived to final node of network at beam %.1f\n", currGenBeam);
         else
            printf("No tokens survived to final node of network\n");
         fflush(stdout);
      } else if (fn==NULL){
         printf("Sorry [%d frames]?\n",nFrames);fflush(stdout);
      }      
      if (pbinfo.a != NULL && replay)  ReplayAudio(pbinfo);
      CloseBuffer(pbuf);
      return FALSE;
   }
   
   if (vri->noTokenSurvived && restartable)
      return FALSE;

   if (vri->noTokenSurvived && trace & T_TOP) {
      printf("No tokens survived to final node of network\n");
      printf("  Output most likely partial hypothesis within network\n");
      fflush(stdout);
   }

   lat->utterance=thisFN;
   lat->net=wdNetFn;
   lat->vocab=dictFn;
   
   if (trace & T_TOP || fn==NULL) {
      node=NULL;
      for (j=0;j<lat->nn;j++) {
         node=lat->lnodes+j;
         if (node->pred==NULL) break;
         node=NULL;
      }
      aclk=lmlk=0.0;
      while(node!=NULL) {
         for (arc=NULL,cur=node->foll;cur!=NULL;cur=cur->farc) arc=cur;
         if (arc==NULL) break;
         if (arc->end->word!=NULL)
            printf("%s ",arc->end->word->wordName->name);
         aclk+=arc->aclike+arc->prlike*lat->prscale;
         lmlk+=arc->lmlike*lat->lmscale+lat->wdpenalty;
         node=arc->end;
      }
      printf(" ==  [%d frames] %.4f [Ac=%.1f LM=%.1f] (Act=%.1f)\n",nFrames,
             (aclk+lmlk)/nFrames, aclk,lmlk,(float)tact/nFrames);
      fflush(stdout);
   }
   if (pbinfo.a != NULL && replay)  ReplayAudio(pbinfo);
   
   /* accumulate stats for online unsupervised adaptation 
      only if a token survived */
   if ((lat != NULL) &&  (!vri->noTokenSurvived) && ((update > 0) || (xfInfo.useOutXForm)))
      DoOnlineAdaptation(lat, pbuf, nFrames);

   if (enableOutput){
      if (nToks>1 && latExt!=NULL) {
         MakeFN(thisFN,labDir,latExt,lfn);
         if ((file=FOpen(lfn,NetOFilter,&isPipe))==NULL) 
            HError(3211,"ProcessFile: Could not open file %s for lattice output",lfn);
         if (latForm==NULL)
            form=HLAT_DEFAULT;
         else {
            for (p=latForm,form=0;*p!=0;p++) {
               switch (*p) {
               case 'A': form|=HLAT_ALABS; break;
               case 'B': form|=HLAT_LBIN; break;
               case 't': form|=HLAT_TIMES; break;
               case 'v': form|=HLAT_PRON; break;
               case 'a': form|=HLAT_ACLIKE; break;
               case 'l': form|=HLAT_LMLIKE; break;
               case 'd': form|=HLAT_ALIGN; break;
               case 'm': form|=HLAT_ALDUR; break;
               case 'n': form|=HLAT_ALLIKE; break;
               case 'r': form|=HLAT_PRLIKE; break;
               }
            }
         }
         if(WriteLattice(lat,file,form)<SUCCESS)
            HError(3214,"ProcessFile: WriteLattice failed");

         FClose(file,isPipe);
      }

      trans=TranscriptionFromLattice(&ansHeap,lat,nTrans);
      
      if (labForm!=NULL)
         FormatTranscription(trans,pbinfo.tgtSampRate,states,models,
                             strchr(labForm,'X')!=NULL,
                             strchr(labForm,'N')!=NULL,strchr(labForm,'S')!=NULL,
                             strchr(labForm,'C')!=NULL,strchr(labForm,'T')!=NULL,
                             strchr(labForm,'W')!=NULL,strchr(labForm,'M')!=NULL);

      MakeFN(thisFN,labDir,labExt,lfn);
      /* if(LSave(lfn,trans,ofmt)<SUCCESS)
         HError(3214,"ProcessFile: Cannot save file %s", lfn); */
      LSave(lfn,trans,ofmt);
      Dispose(&ansHeap,trans);
   }
   Dispose(&ansHeap,lat);
   CloseBuffer(pbuf);
   if (trace & T_MMU){
      printf("Memory State after utter %d\n",utterNum);
      PrintAllHeapStats();
   }

   return !vri->noTokenSurvived;
}

/* --------------------- Top Level Processing --------------------- */

/* DoAlignment: by creating network from transcriptions or lattices */
void DoAlignment(void)
{
   FILE *nf;
   char lfn[255];
   Transcription *trans;
   Network *net;
   Boolean isPipe;
   int n=0;
   LogDouble currGenBeam;
   AdaptXForm *incXForm;

   if (trace&T_TOP) {
      if (loadNetworks) 
         printf("New network will be used for each file\n");
      else
         printf("Label file will be used to align each file\n");
      fflush(stdout);
   }
   CreateHeap(&netHeap,"Net heap",MSTAK,1,0,8000,80000);
   while (NumArgs()>0) {
      if (NextArg() != STRINGARG)
         HError(3219,"DoAlignment: Data file name expected");
      datFN = GetStrArg();
      if (trace&T_TOP) {
         printf("Aligning File: %s\n",datFN);  fflush(stdout);
      }
      MakeFN(datFN,labInDir,labInExt,lfn);
      if (loadNetworks) {
         if ( (nf = FOpen(lfn,NetFilter,&isPipe)) == NULL)
            HError(3210,"DoAlignment: Cannot open Word Net file %s",lfn);
         if((wdNet = ReadLattice(nf,&netHeap,&vocab,TRUE,FALSE))==NULL)
            HError(3210,"DoAlignment: ReadLattice failed");
         FClose(nf,isPipe);
         if (trace&T_TOP) {
            printf("Read lattice with %d nodes / %d arcs\n",
                   wdNet->nn,wdNet->na);
            fflush(stdout);
         }
      }
      else {
         LabList *ll = NULL;

         trans=LOpen(&netHeap,lfn,ifmt);
         if (trans->numLists >= 1)
            ll = GetLabelList(trans,1);
         if (!ll && !bndId)
            HError(3233, "DoAlignment: cannot align empty transcription");

         wdNet=LatticeFromLabels(ll, bndId, &vocab,&netHeap);
         if (trace&T_TOP) {
            printf("Created lattice with %d nodes / %d arcs from label file\n",
                   wdNet->nn,wdNet->na);
            fflush(stdout);
         }
      }
      net=ExpandWordNet(&netHeap,wdNet,&vocab,&hset);

      ++n;
      currGenBeam = genBeam;
      /* This handles the initial input transform, parent transform setting
	 and output transform creation */
      if (UpdateSpkrStats(&hset, &xfInfo, datFN) && (!(xfInfo.useInXForm))) {
         xfInfo.inXForm = NULL;
      }
      if (genBeamInc == 0.0)
         ProcessFile (datFN, net, n, currGenBeam, FALSE);
      else {
         Boolean completed;

         completed = ProcessFile (datFN, net, n, currGenBeam, TRUE);
         currGenBeam += genBeamInc;
         while (!completed && (currGenBeam <= genBeamLim - genBeamInc)) {
            completed = ProcessFile (datFN, net, n, currGenBeam, TRUE);
            currGenBeam += genBeamInc;
         }
         if (!completed)
            ProcessFile (datFN, net, n, currGenBeam, FALSE);
      }

      if (update > 0 && n%update == 0) {
         if (trace&T_TOP) {
            printf("Transforming model set\n");
            fflush(stdout);
         }
	 /* 
	    at every stage a new transform is created - fix?? 
	    Estimate transform and then set it up as the 
	    input XForm
	 */
	 incXForm = CreateAdaptXForm(&hset,"inc");
         TidyBaseAccs();
	 GenAdaptXForm(&hset,incXForm);
         xfInfo.inXForm = GetMLLRDiagCov(incXForm);;
	 SetXForm(&hset,xfInfo.inXForm);
	 ApplyHMMSetXForm(&hset,xfInfo.inXForm);
      }
      ResetHeap(&netHeap);
   }
}

/* DoRecognition:  use single network to recognise each input utterance */
void DoRecognition(void)
{
   FILE *nf;
   Network *net;
   Boolean isPipe;
   int n=0;
   AdaptXForm *incXForm;

   if ( (nf = FOpen(wdNetFn,NetFilter,&isPipe)) == NULL)
      HError(3210,"DoRecognition: Cannot open Word Net file %s",wdNetFn);
   if((wdNet = ReadLattice(nf,&ansHeap,&vocab,TRUE,FALSE))==NULL)
      HError(3210,"DoAlignment: ReadLattice failed");
   FClose(nf,isPipe);

   if (trace&T_TOP) {
      printf("Read lattice with %d nodes / %d arcs\n",wdNet->nn,wdNet->na);
      fflush(stdout);
   }
   CreateHeap(&netHeap,"Net heap",MSTAK,1,0,
              wdNet->na*sizeof(NetLink),wdNet->na*sizeof(NetLink));

   net = ExpandWordNet(&netHeap,wdNet,&vocab,&hset);
   ResetHeap(&ansHeap);
   if (trace&T_TOP) {
      printf("Created network with %d nodes / %d links\n",
             net->numNode,net->numLink);  fflush(stdout);
   }
   if (trace & T_MEM){
      printf("Memory State Before Recognition\n");
      PrintAllHeapStats();
   }

   if (NumArgs()==0) {      /* Process audio */
      while(TRUE){
         printf("\nREADY[%d]>\n",++n); fflush(stdout);
	 /* no input transform possible for audio input .... */
         ProcessFile(NULL,net,n,genBeam, FALSE);
         if (update > 0 && n%update == 0) {
            if (trace&T_TOP) {
               printf("Transforming model set\n");
               fflush(stdout);
            }
	    /* 
	       at every stage a new transform is created - fix?? 
	       Estimate transform and then set it up as the 
	       input XForm
	    */
	    incXForm = CreateAdaptXForm(&hset,"inc");
            TidyBaseAccs();
	    GenAdaptXForm(&hset,incXForm);
            xfInfo.inXForm = GetMLLRDiagCov(incXForm);;
            SetXForm(&hset,xfInfo.inXForm);
	    ApplyHMMSetXForm(&hset,xfInfo.inXForm);
         }
      }
   }
   else {                   /* Process files */
      while (NumArgs()>0) {
         if (NextArg()!=STRINGARG)
            HError(3219,"DoRecognition: Data file name expected");
         datFN = GetStrArg();
         if (trace&T_TOP) {
            printf("File: %s\n",datFN); fflush(stdout);
         }
	 /* This handles the initial input transform, parent transform setting
	    and output transform creation */
         if (UpdateSpkrStats(&hset, &xfInfo, datFN) && (!(xfInfo.useInXForm))) {
            xfInfo.inXForm = NULL;
         }
         ProcessFile(datFN,net,n++,genBeam,FALSE);
         if (update > 0 && n%update == 0) {
            if (trace&T_TOP) {
               printf("Transforming model set\n");
               fflush(stdout);
            }
	    /* 
	       at every stage a new transform is created - fix?? 
	       Estimate transform and then set it up as the 
	       input XForm
	    */
	    incXForm = CreateAdaptXForm(&hset,"inc");
            TidyBaseAccs();
	    GenAdaptXForm(&hset,incXForm);
            xfInfo.inXForm = GetMLLRDiagCov(incXForm);;
            SetXForm(&hset,xfInfo.inXForm);
	    ApplyHMMSetXForm(&hset,xfInfo.inXForm);
         }
      }
   }
}

/* ----------------------------------------------------------- */
/*                      END:  HVite.c                          */
/* ----------------------------------------------------------- */