/* 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;
}
void LoadNetwork()
{
FILE *nf;
Boolean isPipe;
int n=0;
CreateHeap(&wdNetHeap,"Lattice heap",MSTAK,1,0.0,4000,4000);
if ( (nf = FOpen(wdNetFn,NetFilter,&isPipe)) == NULL)
HError(3210,"LoadNetwork: Cannot open Word Net file %s",wdNetFn);
if((wdNet = ReadLattice(nf,&wdNetHeap,&vocab,TRUE,FALSE))==NULL)
HError(3210,"LoadNetwork: ReadLattice failed");
FClose(nf,isPipe);
printf("Read Word Network with %d nodes / %d arcs\n",wdNet->nn,wdNet->na);
CreateHeap(&netHeap,"Net heap",MSTAK,1,0,
wdNet->na*sizeof(NetLink),wdNet->na*sizeof(NetLink));
net = ExpandWordNet(&netHeap,wdNet,&vocab,&hset);
printf("Created network with %d nodes / %d links\n",
net->numNode,net->numLink);
}
/* 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)) && (hset.semiTied == NULL)) {
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);
}
}
}
}
/* DoAlignment: by creating network from transcriptions or lattices */
void DoAlignment(void)
{
FILE *nf;
char lfn[MAXSTRLEN], buf[MAXSTRLEN];
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);
}
if (labFileMask != NULL ) { /* support for rescoring lattice masks */
if (!MaskMatch(labFileMask,buf,datFN))
HError(2319,"DoAlignment: mask %s has no match with segemnt %s",labFileMask,datFN);
MakeFN(buf,labInDir,labInExt,lfn);
} else {
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)) && (hset.semiTied == NULL)) {
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);
}
}