/* LoadFile: load whole file or segments and accumulate variance */
static void LoadFile(char *fn)
{
ParmBuf pbuf;
BufferInfo info;
char labfn[80];
Transcription *trans;
long segStIdx,segEnIdx;
int i,j,ncas,nObs;
LLink p;
if (segId == NULL) { /* load whole parameter file */
if((pbuf=OpenBuffer(&iStack, fn, 0, dff, FALSE_dup, FALSE_dup))==NULL)
HError(2050,"LoadFile: Config parameters invalid");
GetBufferInfo(pbuf,&info);
CheckData(fn,info);
nObs = ObsInBuffer(pbuf);
for (i=0; i<nObs; i++){
ReadAsTable(pbuf,i,&obs);
AccVar(obs);
}
if (trace&T_LOAD) {
printf(" %d observations loaded from %s\n",nObs,fn);
fflush(stdout);
}
CloseBuffer(pbuf);
}
else { /* load segment of parameter file */
MakeFN(fn,labDir,labExt,labfn);
trans = LOpen(&iStack,labfn,lff);
ncas = NumCases(trans->head,segId);
if ( ncas > 0) {
if((pbuf=OpenBuffer(&iStack, fn, 0, dff, FALSE_dup, FALSE_dup))==NULL)
HError(2050,"LoadFile: Config parameters invalid");
GetBufferInfo(pbuf,&info);
CheckData(fn,info);
for (i=1,nObs=0; i<=ncas; i++) {
p = GetCase(trans->head,segId,i);
segStIdx= (long) (p->start/info.tgtSampRate);
segEnIdx = (long) (p->end/info.tgtSampRate);
if (trace&T_SEGS)
printf(" loading seg %s [%ld->%ld]\n",
segId->name,segStIdx,segEnIdx);
if (segEnIdx >= ObsInBuffer(pbuf))
segEnIdx = ObsInBuffer(pbuf)-1;
if (segEnIdx >= segStIdx) {
for (j=segStIdx;j<=segEnIdx;j++) {
ReadAsTable(pbuf,j,&obs);
AccVar(obs); ++nObs;
}
}
}
CloseBuffer(pbuf);
if (trace&T_LOAD)
printf(" %d observations loaded from %s\n",nObs,fn);
}
}
ResetHeap(&iStack);
}
/* LoadTransLabs: Load transcription from file */
Transcription *LoadTransLabs(char *src)
{
Transcription *t;
MakeFN(src,labDir,labExt,labFile);
if(trace & T_SEGMENT)
printf("Loading label file %s\n",labFile);
t = LOpen(&lStack,labFile,srcLabFF);
if(chopF && ! stenSet) SetLabSeg(t);
return t;
}
/* DoGeneration: Generate parameter sequences from HMMs */
void DoGeneration(char *labfn)
{
char labFn[MAXFNAMELEN], buf[MAXSTRLEN];
int t;
Boolean eSep;
Transcription *tr;
if (trace & T_TOP) {
printf(" Generating Label %s\n", NameOf(labfn, buf));
fflush(stdout);
}
/* load a given input label file */
ResetHeap(&utt->transStack);
MakeFN(labfn, labDir, labExt, labFn);
tr = LOpen(&genStack, labFn, lff);
/* compose a sentence HMM corresponding to the input label */
InitialiseGenInfo(genInfo, tr, FALSE);
/* set utterance informations for forward-backward algorithm */
SetStreamWidths(hmset.pkind, hmset.vecSize, hmset.swidth, &eSep);
utt->tr = tr;
utt->Q = genInfo->labseqlen;
utt->T = genInfo->tframe;
utt->twoDataFiles = FALSE;
utt->o = (Observation *) New(&gstack, utt->T * sizeof(Observation));
utt->o--;
for (t = 1; t <= utt->T; t++)
utt->o[t] = MakeObservation(&gstack, hmset.swidth, hmset.pkind, FALSE, eSep);
/* parameter generation */
ParamGen(genInfo, utt, fbInfo, type);
/* output state durations and generated parameter sequences */
if (!stateAlign)
WriteStateDurations(labfn, genInfo);
WriteParms(labfn, genInfo);
/* free memory */
Dispose(&gstack, ++utt->o);
ResetGenInfo(genInfo);
/* increment total number of generated frames */
totalT += utt->T;
totalPr += utt->pr;
return;
}
int main(int argc, char *argv[])
{
char * labFn, *listfn, *s;
int i,fidx;
MLFEntry *me = NULL;
Transcription *t;
void InitStats(char *listfn);
void GatherStats(Transcription *t);
void OutputStats(void);
if(InitShell(argc,argv,hlstats_version,hlstats_vc_id)<SUCCESS)
HError(1300,"HLStats: InitShell failed");
InitMem(); InitMath();
InitWave(); InitLabel();
InitLM();
if (!InfoPrinted() && NumArgs() == 0)
ReportUsage();
if (NumArgs() == 0) Exit(0);
SetConfParms();
enterId=GetLabId("!ENTER",TRUE); /* All sentences should or are coerced */
exitId=GetLabId("!EXIT",TRUE); /* to start enterId and end exitId */
nullId=GetLabId("!NULL",TRUE); /* Name for words not in list */
while (NextArg() == SWITCHARG) {
s = GetSwtArg();
if (strlen(s)!=1)
HError(1319,"HLStats: Bad switch %s; must be single letter",s);
switch(s[0]){
case 'b':
doBigram = TRUE;
if (NextArg() != STRINGARG)
HError(1319,"HLStats: Ngram output file name expected");
bigFile = GetStrArg();
break;
case 'c':
doLCount = TRUE;
lCountLimit = GetChkedInt(0,100000,s);
break;
case 'd':
doDurs = TRUE; break;
case 'f':
bigFloor = GetChkedFlt(0.0,1000.0,s);
break;
case 'h':
hSize = GetChkedInt(1,2,s);
break;
case 'l':
doList = TRUE;
if (NextArg() != STRINGARG)
HError(1319,"HLStats: Output label list file name expected");
listFile = GetStrArg();
break;
case 'o':
doBOff = TRUE;
break;
case 'p':
doPCount = TRUE;
pCountLimit = GetChkedInt(0,100000,s);
break;
case 's':
if (NextArg() != STRINGARG)
HError(1319,"HLStats: ENTER label name expected");
enterId=GetLabId(GetStrArg(),TRUE);
if (NextArg() != STRINGARG)
HError(1319,"HLStats: EXIT label name expected");
exitId=GetLabId(GetStrArg(),TRUE);
break;
case 't':
bigThresh = GetChkedInt(0,100,s);
break;
case 'u':
uniFloor = GetChkedFlt(0.0,1000.0,s);
break;
case 'G':
if (NextArg() != STRINGARG)
HError(1319,"HLStats: Input label File format expected");
if((ff = Str2Format(GetStrArg())) == ALIEN)
HError(-1389,"HLStats: Warning ALIEN Label file format set");
break;
case 'I':
if (NextArg() != STRINGARG)
HError(1319,"HLStats: Input MLF file name expected");
LoadMasterFile(GetStrArg());
break;
case 'T':
if (NextArg() != INTARG)
HError(1319,"HLStats: Trace value expected");
trace = GetChkedInt(0,017,s); break;
default:
HError(1319,"HLStats: Unknown switch %s",s);
}
}
if (NextArg()!=STRINGARG)
HError(1319,"HLStats: Label list file name expected");
listfn = GetStrArg();
if (!(doDurs || doBigram || doList || doLCount || doPCount))
HError(1330,"HLStats: Nothing to do!");
InitStats(listfn);
i=0;
while (NumArgs()>0) {
if (NextArg()!=STRINGARG)
HError(1319,"HLStats: Input label file name expected");
labFn = GetStrArg();
if (IsMLFFile(labFn)) {
fidx = NumMLFFiles();
if ((me=GetMLFTable()) != NULL) {
while(me->next != NULL) me=me->next;
LoadMasterFile(labFn);
me=me->next;
}
else {
LoadMasterFile(labFn);
me=GetMLFTable();
}
while (me != NULL) {
if (me->type == MLF_IMMEDIATE && me->def.immed.fidx == fidx) {
if (trace&T_FIL) {
printf(" Processing file %s\n",me->pattern); fflush(stdout);
}
t = LOpen(&tmpHeap,me->pattern,ff);
if (t->numLists<1)
HError(-1330,"HLStats: Empty file %s",me->pattern);
else
GatherStats(t),i++;
Dispose(&tmpHeap,t);
}
me = me->next;
if ((trace&T_BAS) && !(trace&T_FIL) &&
NumMLFEntries()>5000 && i%1000==0)
printf(". "),fflush(stdout);
}
if ((trace&T_BAS) && !(trace&T_FIL) && NumMLFEntries()>5000)
printf("\n");
} else {
if (trace&T_FIL) {
printf(" Processing file %s\n",labFn); fflush(stdout);
}
t = LOpen(&tmpHeap,labFn,ff);
if (t->numLists<1)
HError(-1330,"HLStats: Empty file %s",me->pattern);
else
GatherStats(t),i++;
Dispose(&tmpHeap,t);
}
}
if (trace&T_MEM)
PrintAllHeapStats();
OutputStats();
if (trace&T_MEM)
PrintAllHeapStats();
Exit(0);
return (0); /* never reached -- make compiler happy */
}
/* LoadFile: load whole file or segments into segStore */
void LoadFile(char *fn)
{
BufferInfo info;
char labfn[80];
Transcription *trans;
long segStIdx,segEnIdx;
static int segIdx=1; /* Between call handle on latest seg in segStore */
static int prevSegIdx=1;
HTime tStart, tEnd;
int i,k,s,ncas,nObs=0,segLen;
LLink p;
Observation obs;
if((pbuf=OpenBuffer(&bufferStack, fn, 10, dff, FALSE_dup, FALSE_dup))==NULL)
HError(2150,"LoadFile: Config parameters invalid");
GetBufferInfo(pbuf,&info);
CheckData(fn,info);
if (firstTime) InitSegStore(&info);
if (segId == NULL) { /* load whole parameter file */
nObs = ObsInBuffer(pbuf);
tStart = 0.0;
tEnd = (info.tgtSampRate * nObs);
LoadSegment(segStore, tStart, tEnd, pbuf);
segIdx++;
}
else { /* load segment of parameter file */
MakeFN(fn,labDir,labExt,labfn);
trans = LOpen(&transStack,labfn,lff);
ncas = NumCases(trans->head,segId);
if ( ncas > 0) {
for (i=1,nObs=0; i<=ncas; i++) {
p = GetCase(trans->head,segId,i);
segStIdx = (long)(p->start/info.tgtSampRate);
segEnIdx = (long)(p->end/info.tgtSampRate);
if (segEnIdx >= ObsInBuffer(pbuf))
segEnIdx = ObsInBuffer(pbuf)-1;
if (segEnIdx - segStIdx + 1 >= nStates-2) {
LoadSegment(segStore, p->start, p->end, pbuf);
if (trace&T_LD1)
printf(" loading seg %s %f[%ld]->%f[%ld]\n",segId->name,
p->start,segStIdx,p->end,segEnIdx);
nObs += SegLength(segStore, segIdx);
segIdx++;
}else if (trace&T_LD1)
printf(" seg %s %f->%f ignored\n",segId->name,
p->start,p->end);
}
}
}
if (hset.hsKind == DISCRETEHS){
for (k=prevSegIdx; k<segIdx; k++){
segLen = SegLength(segStore, k);
for (i=1; i<=segLen; i++){
obs = GetSegObs(segStore, k, i);
for (s=1; s<=nStreams; s++){
if( (obs.vq[s] < 1) || (obs.vq[s] > maxMixInS[s]))
HError(2150,"LoadFile: Discrete data value [ %d ] out of range in stream [ %d ] in file %s",obs.vq[s],s,fn);
}
}
}
prevSegIdx=segIdx;
}
if (trace&T_LD0)
printf(" %d observations loaded from %s\n",nObs,fn);
CloseBuffer(pbuf);
ResetHeap(&transStack);
}
/* LoadFile: load whole file or segments and accumulate variance */
void LoadFile(char *fn)
{
ParmBuf pbuf;
BufferInfo info;
char labfn[80];
Transcription *trans;
long segStIdx,segEnIdx;
int i,s,j,ncas,nObs=0;
LLink p;
if (segId == NULL) { /* load whole parameter file */
if((pbuf=OpenBuffer(&iStack, fn, 0, dff, FALSE_dup, FALSE_dup))==NULL)
HError(2550,"LoadFile: Config parameters invalid");
GetBufferInfo(pbuf,&info);
CheckData(fn,info);
nObs = ObsInBuffer(pbuf);
for (i=0; i<nObs; i++) {
for(s=1;s<=swidth[0];s++)
obs.fv[s] = CreateVector(&dStack,swidth[s]);
ReadAsTable(pbuf,i,&obs);
for(s=1;s<=swidth[0];s++)
StoreItem(dSeq[s],(Ptr)obs.fv[s]);
}
CloseBuffer(pbuf);
}
else { /* load segment of parameter file */
MakeFN(fn,labDir,labExt,labfn);
trans = LOpen(&iStack,labfn,lff);
ncas = NumCases(trans->head,segId);
if ( ncas > 0) {
if((pbuf=OpenBuffer(&iStack, fn, 0, dff, FALSE_dup, FALSE_dup))==NULL)
HError(2550,"LoadFile: Config parameters invalid");
GetBufferInfo(pbuf,&info);
CheckData(fn,info);
for (i=1,nObs=0; i<=ncas; i++) {
p = GetCase(trans->head,segId,i);
segStIdx= (long) (p->start/info.tgtSampRate);
segEnIdx = (long) (p->end/info.tgtSampRate);
if (trace&T_SEGS)
printf(" loading seg %s [%ld->%ld]\n",
segId->name,segStIdx,segEnIdx);
if (segEnIdx >= ObsInBuffer(pbuf))
segEnIdx = ObsInBuffer(pbuf)-1;
if (segEnIdx >= segStIdx) {
for (j=segStIdx;j<=segEnIdx;j++) {
/* SJY: The HInit code I copied this from had no */
/* SJY: CreateVector call here -- a bug? */
for(s=1;s<=swidth[0];s++)
obs.fv[s] = CreateVector(&dStack,swidth[s]);
ReadAsTable(pbuf,j,&obs);
for(s=1;s<=swidth[0];s++)
StoreItem(dSeq[s],(Ptr)obs.fv[s]);
++nObs;
}
}
}
CloseBuffer(pbuf);
}
}
ResetHeap(&iStack);
if (trace&T_LOAD) {
printf(" %5d obs loaded from %s, streams: ",nObs,fn);
for(s=1;s<=swidth[0];s++) printf("[%d]" ,swidth[s]);
printf("\n"); fflush(stdout);
}
}
/* ProcessLabelFile: compute perplexity and related statistics from labels */
static void ProcessLabelFile(char *fn, int nSize)
{
LLink ll;
double ppl;
LabList *ref;
LabId lab;
Transcription *tr;
int i,numPLabs,nLabel;
tr = LOpen(&tempHeap, fn, lff);
if (tr->numLists < 1) {
HError(-16635,"ProcessLabelFile: transcription file %s is Empty",fn);
return;
}
ref = GetLabelList(tr, 1);
if (ref->head->succ == ref->tail) {
HError(-16635,"ProcessLabelFile: transcription file %s is Empty",fn);
return;
}
if (trace>0) {
printf("Processing label file: %s\n", fn);
fflush(stdout);
}
nLabel = CountLabs(ref);
ZeroStats(&sent);
sent.nTok = nLabel + 2; sent.nUtt = 1;
/* copy labels into pLab, mapping OOVs */
numPLabs = 0;
if (sstId!=NULL) /* add sentence start marker(s) */
for (i=0; i<(nSize-1); i++) pLab[numPLabs++] = sstId;
for (i=0,ll=ref->head->succ; i<nLabel; i++,ll=ll->succ) {
lab = GetEQLab(ll->labid);
if ((i==0) && IS_SST(lab)) {
sent.nTok--; continue;
}
if ((i==(nLabel-1)) && IS_SEN(lab)) {
sent.nTok--; continue;
}
if (IS_UNK(lab)) {
if (trace&T_OOV)
printf("mapping OOV: %s\n", lab->name);
StoreOOV(&sent,lab,1); lab = unkId;
}
pLab[numPLabs++] = lab;
if (numPLabs>=LBUF_SIZE) {
HError(16650, "Maximum utterance length in a label file exceeded (limit is compiled to be %d tokens)",
LBUF_SIZE);
}
}
if (senId!=NULL) /* add sentence end marker */
pLab[numPLabs++] = senId;
CalcPerplexity(&sent, pLab, numPLabs, nSize);
AddStats(&sent, &totl);
if (trace&T_SEL) { /* compact info for sentence selection */
ppl = exp(-(sent.logpp)/(double) (sent.nWrd));
printf("#! %.4f", ppl);
for (i=0, ll=ref->head->succ; i<nLabel; i++, ll=ll->succ)
printf(" %s", ll->labid->name);
printf("\n"); fflush(stdout);
}
}
/* 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);
}
}