int main(int argc, char *argv[])
{
char *s;
char *labfn;
int numUtt;
void Initialise(void);
void DoGeneration(char *labfn);
if (InitShell(argc, argv, hmgens_version, hmgens_vc_id) < SUCCESS)
HError(2300, "HMGenS: InitShell failed");
InitMem();
InitMath();
InitSigP();
InitWave();
InitLabel();
InitModel();
if (InitParm() < SUCCESS)
HError(2300, "HMGenS: InitParm failed");
InitUtil();
InitFB();
InitAdapt(&xfInfo_hmm, &xfInfo_dur);
InitMap();
InitGen();
if (NumArgs() == 0)
ReportUsage();
CreateHeap(&genStack, "genStore", MSTAK, 1, 1.0, 80000, 400000);
CreateHeap(&uttStack, "uttStore", MSTAK, 1, 0.5, 100, 1000);
CreateHeap(&fbInfoStack, "FBInfoStore", MSTAK, 1, 0.5, 100, 1000);
CreateHeap(&hmmStack, "HmmStore", MSTAK, 1, 1.0, 50000, 500000);
CreateHeap(&dmStack, "dmStore", MSTAK, 1, 1.0, 50000, 500000);
SetConfParms();
CreateHMMSet(&hmset, &hmmStack, TRUE);
CreateHMMSet(&dmset, &dmStack, TRUE);
utt = (UttInfo *) New(&uttStack, sizeof(UttInfo));
genInfo = (GenInfo *) New(&genStack, sizeof(GenInfo));
fbInfo = (FBInfo *) New(&fbInfoStack, sizeof(FBInfo));
while (NextArg() == SWITCHARG) {
s = GetSwtArg();
if (strlen(s) != 1)
HError(9919, "HMGenS: Bad switch %s; must be single letter", s);
switch (s[0]) {
case 'a':
xfInfo_hmm.useInXForm = TRUE;
break;
case 'b':
xfInfo_dur.useInXForm = TRUE;
break;
case 'c':
if (NextArg() != INTARG)
HError(2119, "HMGenS: Parameter generation algorithm type value expected");
type = (ParmGenType) GetChkedInt(CHOLESKY, FB, s);
break;
case 'd':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: HMM definition directory expected");
hmmDir = GetStrArg();
break;
case 'e':
useAlign = TRUE;
break;
case 'f':
frameRate = (HTime) GetChkedFlt(0.0, 10000000.0, s);
break;
case 'g':
minFrwdP = GetChkedFlt(0.0, 1000.0, s);
break;
case 'h':
if (NextArg() != STRINGARG)
HError(1, "Speaker name pattern expected");
xfInfo_hmm.inSpkrPat = xfInfo_dur.inSpkrPat = GetStrArg();
if (NextArg() == STRINGARG)
xfInfo_hmm.paSpkrPat = xfInfo_dur.paSpkrPat = GetStrArg();
if (NextArg() != SWITCHARG)
HError(2319, "HMGenS: cannot have -h as the last option");
break;
case 'm':
modelAlign = TRUE;
break;
case 'n':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: duration model definition directory expected");
dmDir = GetStrArg();
break;
case 'p':
outPdf = TRUE;
break;
case 'r':
if (NextArg() != FLOATARG)
HError(2119, "HMGenS: Speaking rate value (float) expected");
speakRate = GetChkedFlt(0.0, 3.0, s);
break;
case 's':
stateAlign = TRUE;
break;
case 't':
pruneInit = GetChkedFlt(0.0, 1.0E20, s);
if (NextArg() == FLOATARG || NextArg() == INTARG) {
pruneInc = GetChkedFlt(0.0, 1.0E20, s);
pruneLim = GetChkedFlt(0.0, 1.0E20, s);
} else {
pruneInc = 0.0;
pruneLim = pruneInit;
}
break;
case 'v':
MSDthresh = GetChkedFlt(0.0, 1.0, s);
break;
case 'x':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: HMM file extension expected");
hmmExt = GetStrArg();
break;
case 'y':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: duration model file extension expected");
dmExt = GetStrArg();
break;
case 'B':
inBinary = TRUE;
break;
case 'E':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: parent transform directory expected");
xfInfo_hmm.usePaXForm = TRUE;
xfInfo_hmm.paXFormDir = GetStrArg();
if (NextArg() == STRINGARG)
xfInfo_hmm.paXFormExt = GetStrArg();
if (NextArg() != SWITCHARG)
HError(2319, "HMGenS: cannot have -E as the last option");
break;
case 'G':
if (NextArg() != STRINGARG)
HError(2119, "HMGenS: Label File format expected");
if ((lff = Str2Format(GetStrArg())) == ALIEN)
HError(-2189, "HMGenS: Warning ALIEN Label file format set");
break;
case 'H':
if (NextArg() != STRINGARG)
HError(3219, "HMGenS: HMM MMF File name expected");
AddMMF(&hmset, GetStrArg());
break;
case 'I':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: MLF file name expected");
LoadMasterFile(GetStrArg());
break;
case 'J':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: input transform directory expected");
AddInXFormDir(&hmset, GetStrArg());
if (NextArg() == STRINGARG)
xfInfo_hmm.inXFormExt = GetStrArg();
if (NextArg() != SWITCHARG)
HError(2319, "HMGenS: cannot have -J as the last option");
break;
case 'L':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: Label file directory expected");
labDir = GetStrArg();
break;
case 'M':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: Output macro file directory expected");
genDir = GetStrArg();
break;
case 'N':
if (NextArg() != STRINGARG)
HError(3219, "HMGenS: Duration MMF File name expected");
AddMMF(&dmset, GetStrArg());
break;
case 'T':
if (NextArg() != INTARG)
HError(2119, "HMGenS: Trace value expected");
trace = GetChkedInt(0, 0002, s);
break;
case 'W':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: parent duration transform directory expected");
xfInfo_dur.usePaXForm = TRUE;
xfInfo_dur.paXFormDir = GetStrArg();
if (NextArg() == STRINGARG)
xfInfo_dur.paXFormExt = GetStrArg();
if (NextArg() != SWITCHARG)
HError(2319, "HMGenS: cannot have -W as the last option");
break;
case 'X':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: Label file extension expected");
labExt = GetStrArg();
break;
case 'Y':
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: input duration transform directory expected");
AddInXFormDir(&dmset, GetStrArg());
if (NextArg() == STRINGARG) {
if (xfInfo_dur.inXFormExt == NULL)
xfInfo_dur.inXFormExt = GetStrArg();
else
HError(2319, "MGenS: only one input duration transform extension may be specified");
}
if (NextArg() != SWITCHARG)
HError(2319, "HMGenS: cannot have -Y as the last option");
break;
default:
HError(9919, "HMGenS: Unknown switch %s", s);
}
}
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: file name of vocabulary list expected");
Initialise();
InitUttInfo(utt, FALSE);
numUtt = 1;
if (trace & T_TOP)
SetTraceGen();
/* generate parameter sequences */
do {
if (NextArg() != STRINGARG)
HError(2319, "HMGenS: data file name expected");
labfn = GetStrArg();
/* track speakers */
if (UpdateSpkrStats(&hmset, &xfInfo_hmm, labfn)) {
if (!xfInfo_hmm.useInXForm)
xfInfo_hmm.inXForm = NULL;
}
if (UpdateSpkrStats(&dmset, &xfInfo_dur, labfn)) {
if (!xfInfo_dur.useInXForm)
xfInfo_dur.inXForm = NULL;
else
ResetDMMPreComps(&dmset);
}
fbInfo->xfinfo_hmm = &xfInfo_hmm;
fbInfo->xfinfo_dur = &xfInfo_dur;
fbInfo->inXForm_hmm = xfInfo_hmm.inXForm;
fbInfo->inXForm_dur = xfInfo_dur.inXForm;
fbInfo->al_inXForm_hmm = xfInfo_hmm.al_inXForm;
fbInfo->al_inXForm_dur = xfInfo_dur.al_inXForm;
fbInfo->paXForm_hmm = xfInfo_hmm.paXForm;
fbInfo->paXForm_dur = xfInfo_dur.paXForm;
/* generate parameters */
DoGeneration(labfn);
numUtt++;
} while (NumArgs() > 0);
if (trace & T_TOP) {
printf("Generation complete - average log prob per frame = %e (%d frames)\n", totalPr / totalT, totalT);
}
/* Reset stacks */
Dispose(&fbInfoStack, fbInfo);
Dispose(&genStack, genInfo);
Dispose(&uttStack, utt);
ResetHeap(&fbInfoStack);
ResetHeap(&uttStack);
ResetHeap(&genStack);
ResetHeap(&dmStack);
ResetHeap(&hmmStack);
/* Reset modules */
ResetGen();
ResetAdapt(&xfInfo_hmm, &xfInfo_dur);
ResetFB();
ResetUtil();
ResetParm();
ResetModel();
ResetLabel();
ResetWave();
ResetSigP();
ResetMath();
ResetMem();
ResetShell();
Exit(0);
return (0); /* never reached -- make compiler happy */
}
/* SetConfParms: set conf parms relevant to HCompV */
void SetConfParms(void)
{
int i;
Boolean b;
double f;
char buf[MAXSTRLEN];
nParm = GetConfig("HEREST", TRUE, cParm, MAXGLOBS);
if (nParm>0) {
if (GetConfInt(cParm,nParm,"TRACE",&i)) trace = i;
if (GetConfFlt(cParm,nParm,"VARFLOORPERCENTILE",&f)) varFloorPercent = f;
if (GetConfBool(cParm,nParm,"SAVEBINARY",&b)) saveBinary = b;
if (GetConfBool(cParm,nParm,"BINARYACCFORMAT",&b)) ldBinary = b;
/* 2-model reestimation alignment model set */
if (GetConfStr(cParm,nParm,"ALIGNMODELMMF",buf)) {
strcpy(al_hmmMMF,buf); al_hmmUsed = TRUE;
}
if (GetConfStr(cParm,nParm,"ALIGNHMMLIST",buf)) {
strcpy(al_hmmLst,buf); al_hmmUsed = TRUE;
}
/* allow multiple individual model files */
if (GetConfStr(cParm,nParm,"ALIGNMODELDIR",buf)) {
strcpy(al_hmmDir,buf); al_hmmUsed = TRUE;
}
if (GetConfStr(cParm,nParm,"ALIGNMODELEXT",buf)) {
strcpy(al_hmmExt,buf); al_hmmUsed = TRUE;
}
if (GetConfStr(cParm,nParm,"ALIGNXFORMEXT",buf)) {
xfInfo.alXFormExt = CopyString(&hmmStack,buf);
}
if (GetConfStr(cParm,nParm,"ALIGNXFORMDIR",buf)) {
xfInfo.alXFormDir = CopyString(&hmmStack,buf);
}
if (GetConfStr(cParm,nParm,"INXFORMMASK",buf)) {
xfInfo.inSpkrPat = CopyString(&hmmStack,buf);
}
if (GetConfStr(cParm,nParm,"PAXFORMMASK",buf)) {
xfInfo.paSpkrPat = CopyString(&hmmStack,buf);
}
if (GetConfStr(cParm,nParm,"LABFILEMASK",buf)) {
labFileMask = (char*)malloc(strlen(buf)+1);
strcpy(labFileMask, buf);
}
if (GetConfStr(cParm,nParm,"UPDATEMODE",buf)) {
if (!strcmp (buf, "DUMP")) updateMode = UPMODE_DUMP;
else if (!strcmp (buf, "UPDATE")) updateMode = UPMODE_UPDATE;
else if (!strcmp (buf, "BOTH")) updateMode = UPMODE_BOTH;
else HError(2319, "Unknown UPDATEMODE specified (must be DUMP, UPDATE or BOTH)");
}
}
}
void ReportUsage(void)
{
printf("\nUSAGE: HERest [options] hmmList dataFiles...\n\n");
printf(" Option Default\n\n");
printf(" -a Use an input linear transform off\n");
printf(" -c f Mixture pruning threshold 10.0\n");
printf(" -d s dir to find hmm definitions current\n");
printf(" -h s set output speaker name pattern *.%%%%%%\n");
printf(" to s, optionally set input and parent patterns\n");
printf(" -l N set max files per speaker off\n");
printf(" -m N set min examples needed per model 3\n");
printf(" -o s extension for new hmm files as src\n");
printf(" -p N set parallel mode to N off\n");
printf(" -r Enable Single Pass Training... \n");
printf(" ...using two parameterisations off\n");
printf(" -s s print statistics to file s off\n");
printf(" -t f [i l] set pruning to f [inc limit] inf\n");
printf(" -u tmvwap update t)rans m)eans v)ars w)ghts tmvw\n");
printf(" a)daptation xform p)rior used \n");
printf(" s)semi-tied xform \n");
printf(" -v f set minimum variance to f 0.0\n");
printf(" -w f set mix weight floor to f*MINMIX 0.0\n");
printf(" -x s extension for hmm files none\n");
printf(" -z s Save all xforms to TMF file s TMF\n");
PrintStdOpts("BEFGHIJKLMSTX");
printf("\n\n");
}
void SetuFlags(void)
{
char *s;
s=GetStrArg();
uFlags=(UPDSet) 0;
while (*s != '\0')
switch (*s++) {
case 't': uFlags = (UPDSet) (uFlags+UPTRANS); break;
case 'm': uFlags = (UPDSet) (uFlags+UPMEANS); break;
case 'v': uFlags = (UPDSet) (uFlags+UPVARS); break;
case 'w': uFlags = (UPDSet) (uFlags+UPMIXES); break;
case 's': uFlags = (UPDSet) (uFlags+UPSEMIT); break;
case 'a': uFlags = (UPDSet) (uFlags+UPXFORM); break;
case 'p': uFlags = (UPDSet) (uFlags+UPMAP); break;
default: HError(2320,"SetuFlags: Unknown update flag %c",*s);
break;
}
}
/* ScriptWord: return next word from script */
char *ScriptWord(FILE *script, char *scriptBuf)
{
int ch,qch,i;
i=0; ch=' ';
while (isspace(ch)) ch = fgetc(script);
if (ch==EOF) {
scriptBuf=NULL;
return NULL;
}
if (ch=='\'' || ch=='"'){
qch = ch;
ch = fgetc(script);
while (ch != qch && ch != EOF) {
scriptBuf[i++] = ch;
ch = fgetc(script);
}
if (ch==EOF)
HError(5051,"ScriptWord: Closing quote missing in script file");
} else {
do {
scriptBuf[i++] = ch;
ch = fgetc(script);
}while (!isspace(ch) && ch != EOF);
}
scriptBuf[i] = '\0';
return scriptBuf;
}
void CheckUpdateSetUp()
{
AdaptXForm *xf;
/* Initialise: set up global data structures */
void Initialise(void)
{
Boolean eSep;
int s;
/* Load hmms, convert to inverse DiagC */
if(MakeHMMSet(&hset,hmmListFn)<SUCCESS)
HError(3228,"Initialise: MakeHMMSet failed");
if(LoadHMMSet(&hset,hmmDir,hmmExt)<SUCCESS)
HError(3228,"Initialise: LoadHMMSet failed");
ConvDiagC(&hset,TRUE);
/* Create observation and storage for input buffer */
SetStreamWidths(hset.pkind,hset.vecSize,hset.swidth,&eSep);
obs=MakeObservation(&gstack,hset.swidth,hset.pkind,
hset.hsKind==DISCRETEHS,eSep);
/* sort out masks just in case using adaptation */
if (xfInfo.inSpkrPat == NULL) xfInfo.inSpkrPat = xfInfo.outSpkrPat;
if (xfInfo.paSpkrPat == NULL) xfInfo.paSpkrPat = xfInfo.outSpkrPat;
if (xfInfo.useOutXForm || (update>0)) {
CreateHeap(®Heap, "regClassStore", MSTAK, 1, 0.5, 1000, 8000 );
/* This initialises things - temporary hack - THINK!! */
CreateAdaptXForm(&hset, "tmp");
/* initialise structures for the f-b frame-state alignment pass */
utt = (UttInfo *) New(®Heap, sizeof(UttInfo));
fbInfo = (FBInfo *) New(®Heap, sizeof(FBInfo));
/* initialise a recogniser for frame/state alignment purposes */
alignpsi=InitPSetInfo(&hset);
alignvri=InitVRecInfo(alignpsi,1,TRUE,FALSE);
SetPruningLevels(alignvri,0,genBeam,-LZERO,0.0,tmBeam);
InitUttInfo(utt, FALSE);
InitialiseForBack(fbInfo, ®Heap, &hset,
(UPDSet) (UPXFORM), genBeam*2.0, genBeam*2.0,
genBeam*4.0+1.0, 10.0);
utt->twoDataFiles = FALSE;
utt->S = hset.swidth[0];
AttachPreComps(&hset,hset.hmem);
}
CreateHeap(&bufHeap,"Input Buffer heap",MSTAK,1,0.0,50000,50000);
CreateHeap(&repHeap,"Replay Buffer heap",MSTAK,1,0.0,50000,50000);
maxM = MaxMixInSet(&hset);
for (s=1; s<=hset.swidth[0]; s++)
maxMixInS[s] = MaxMixInSetS(&hset, s);
if (trace&T_TOP) {
printf("Read %d physical / %d logical HMMs\n",
hset.numPhyHMM,hset.numLogHMM); fflush(stdout);
}
/* Initialise recogniser */
if (nToks>1) nBeam=genBeam;
psi=InitPSetInfo(&hset);
vri=InitVRecInfo(psi,nToks,models,states);
/* Read dictionary and create storage for lattice */
InitVocab(&vocab);
if(ReadDict(dictFn,&vocab)<SUCCESS)
HError(3213, "Main: ReadDict failed");
CreateHeap(&ansHeap,"Lattice heap",MSTAK,1,0.0,4000,4000);
if (trace & T_MEM){
printf("Memory State After Initialisation\n");
PrintAllHeapStats();
}
}