/* DoOnlineAdaptation: Perform unsupervised online adaptation
using the recognition hypothesis as the transcription */
int DoOnlineAdaptation(Lattice *lat, ParmBuf pbuf, int nFrames)
{
Transcription *modelTrans, *trans;
BufferInfo pbinfo;
Lattice *alignLat, *wordNet;
Network *alignNet;
int i;
GetBufferInfo(pbuf,&pbinfo);
trans=TranscriptionFromLattice(&netHeap,lat,1);
wordNet=LatticeFromLabels(GetLabelList(trans,1),bndId,
&vocab,&netHeap);
alignNet=ExpandWordNet(&netHeap,wordNet,&vocab,&hset);
StartRecognition(alignvri,alignNet,0.0,0.0,0.0);
/* do forced alignment */
for (i = 0; i < nFrames; i++) {
ReadAsTable(pbuf, i, &obs);
ProcessObservation(alignvri,&obs,-1,xfInfo.inXForm);
}
alignLat=CompleteRecognition(alignvri,
pbinfo.tgtSampRate/10000000.0,
&netHeap);
if (alignvri->noTokenSurvived) {
Dispose(&netHeap, trans);
/* Return value 0 to indicate zero frames process failed */
return 0;
}
modelTrans=TranscriptionFromLattice(&netHeap,alignLat,1);
/* format the transcription so that it contains just the models */
FormatTranscription(modelTrans,pbinfo.tgtSampRate,FALSE,TRUE,
FALSE,FALSE,TRUE,FALSE,TRUE,TRUE, FALSE);
/* Now do the frame/state alignment accumulating MLLR statistics */
/* set the various values in the utterance storage */
utt->tr = modelTrans;
utt->pbuf = pbuf;
utt->Q = CountLabs(utt->tr->head);
utt->T = nFrames;
utt->ot = obs;
/* do frame state alignment and accumulate statistics */
fbInfo->inXForm = xfInfo.inXForm;
fbInfo->al_inXForm = xfInfo.inXForm;
fbInfo->paXForm = xfInfo.paXForm;
if (!FBFile(fbInfo, utt, NULL))
nFrames = 0;
Dispose(&netHeap, trans);
if (trace&T_TOP) {
printf("Accumulated statistics...\n");
fflush(stdout);
}
return nFrames;
}
// scan the lattice and generate output packets
void ARec::ComputeAnswer()
{
int i;
PartialPath pp;
float score;
Lattice *lat, *prunedlat;
Transcription *trans;
pp = FinalBestPath(pri);
// PrintLatStuff(pp);
if (pp.n==0) pp = CurrentBestPath(pri);
if (trace&T_FAN){
printf("Final complete answer\n");
PrintPartialPath(pp,TRUE);
printf("-----------\n");
}
// simple 1 best trace for now - need to make this into a chart
if (runmode&RESULT_ATEND) {
if(nBest>0){
lat=NULL;
lat=CreateLatticeFromOutput(pri, pp, &ansHeap, sampPeriod);
//prunedlat=LatPrune(&ansHeap, lat, 300, 30);
if(lat!=NULL) {
trans=TransFromLattice(&altHeap, lat, nBest);
if(trace&T_OUT)
PrintTranscription(trans,"Output Transcription");
/*DoConfGen(&ansHeap, lat);*/
OutTranscription(trans);
Dispose(&ansHeap,lat);
}
} else {
for (i=1; i<=pp.n; i++)
OutPathElement(i,pp);
}
}
// Send an END packet
if (runmode&RESULT_ALL) {
score = 0.0;
if (pp.n>0) {
assert(pp.path);
score = float(pp.path->like);
}
OutPacket(End_PT, "", "",outseqnum,0,0.0,0.0,score,-1,
(float)tact/frameCount,enTime,enTime);
}
CompleteRecognition(pri);
OutMarkers(-1);
}
/* ProcessFile: process given file. If fn=NULL then direct audio */
Boolean ProcessFile(char *fn, Network *net, int utterNum, LogDouble currGenBeam, Boolean restartable)
{
FILE *file;
ParmBuf pbuf;
BufferInfo pbinfo;
NetNode *d;
Lattice *lat;
LArc *arc,*cur;
LNode *node;
Transcription *trans;
MLink m;
LogFloat lmlk,aclk;
int s,j,tact,nFrames;
LatFormat form;
char *p,lfn[255],buf1[80],buf2[80],thisFN[MAXSTRLEN];
Boolean enableOutput = TRUE, isPipe;
if (fn!=NULL)
strcpy(thisFN,fn);
else if (fn==NULL && saveAudioOut)
CounterFN(roPrefix,roSuffix,++roCounter,4,thisFN);
else
enableOutput = FALSE;
if((pbuf = OpenBuffer(&bufHeap,fn,50,dfmt,TRI_UNDEF,TRI_UNDEF))==NULL)
HError(3250,"ProcessFile: Config parameters invalid");
/* Check pbuf same as hset */
GetBufferInfo(pbuf,&pbinfo);
if (pbinfo.tgtPK!=hset.pkind)
HError(3231,"ProcessFile: Incompatible sample kind %s vs %s",
ParmKind2Str(pbinfo.tgtPK,buf1),
ParmKind2Str(hset.pkind,buf2));
if (pbinfo.a != NULL && replay) AttachReplayBuf(pbinfo.a, (int) (3*(1.0E+07/pbinfo.srcSampRate)));
StartRecognition(vri,net,lmScale,wordPen,prScale);
SetPruningLevels(vri,maxActive,currGenBeam,wordBeam,nBeam,tmBeam);
tact=0;nFrames=0;
StartBuffer(pbuf);
while(BufferStatus(pbuf)!=PB_CLEARED) {
ReadAsBuffer(pbuf,&obs);
if (trace&T_OBS) PrintObservation(nFrames,&obs,13);
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);
}
/* only output 1-best transcription if generating lattices */
if (nTrans > 1 && latExt != NULL)
trans=TranscriptionFromLattice(&ansHeap,lat,1);
/* output N-best transcriptions as usual */
else
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;
}
