hmodel.c

该压缩包为最新版htk的源代码,htk是现在比较流行的语音处理软件,请有兴趣的朋友下载使用

/* ----------------------------------------------------------- */
/*                                                             */
/*                          ___                                */
/*                       |_| | |_/   SPEECH                    */
/*                       | | | | \   RECOGNITION               */
/*                       =========   SOFTWARE                  */ 
/*                                                             */
/*                                                             */
/* ----------------------------------------------------------- */
/* developed at:                                               */
/*                                                             */
/*      Speech Vision and Robotics group                       */
/*      Cambridge University Engineering Department            */
/*      http://svr-www.eng.cam.ac.uk/                          */
/*                                                             */
/*      Entropic Cambridge Research Laboratory                 */
/*      (now part of Microsoft)                                */
/*                                                             */
/* ----------------------------------------------------------- */
/*         Copyright: Microsoft Corporation                    */
/*          1995-2000 Redmond, Washington USA                  */
/*                    http://www.microsoft.com                 */
/*                                                             */
/*               2002 Cambridge University                     */
/*                    Engineering Department                   */
/*                                                             */
/*   Use of this software is governed by a License Agreement   */
/*    ** See the file License for the Conditions of Use  **    */
/*    **     This banner notice must not be removed      **    */
/*                                                             */
/* ----------------------------------------------------------- */
/*         File: HModel.c  HMM Model Definition Data Type      */
/* ----------------------------------------------------------- */

char *hmodel_version = "!HVER!HModel:   3.3 [CUED 28/04/05]";
char *hmodel_vc_id = "$Id: HModel.c,v 1.4 2005/07/22 10:17:01 mjfg Exp $";

#include "HShell.h"
#include "HMem.h"
#include "HMath.h"
#include "HWave.h"
#include "HAudio.h"
#include "HParm.h"
#include "HLabel.h"
#include "HModel.h"
#include "HUtil.h"
#include "HTrain.h"
#include "HAdapt.h"

/* --------------------------- Trace Flags ------------------------- */

static int trace = 0;

#define T_TOP  00001       /* Top Level tracing */
#define T_CHK  00002       /* Show HMM Checking */
#define T_TOK  00004       /* Scanner Token Processing */
#define T_PAR  00010       /* Trace Parsing */
#define T_PMP  00020       /* Pointer Map Tracing */
#define T_MAC  00040       /* Macro Load/Save Tracing */
#define T_ORP  00100       /* Orphan macro handling */
#define T_BTR  00200       /* Decision Tree handling */
#define T_GMX  00400       /* GMP optimisation */
#define T_XFM  01000       /* Loading of xform macros */
#define T_XFD  02000       /* Additional detail of loading of xform macros */

#define CREATEFIDX -1
#define LOADFIDX   -2

/* ------------------ Input XForm directory info ------------------- */

typedef struct _XFDirInfo *XFDirLink;

typedef struct _XFDirInfo {
  char *dirName;           /* input XForm directory name */
  XFDirLink next;          /* next directory name in list */
} XFDirInfo;

/* --------------------------- Initialisation ---------------------- */

static ConfParam *cParm[MAXGLOBS];      /* config parameters */
static int nParm = 0;
static Boolean checking   = TRUE;       /* check HMM defs */
static Boolean saveBinary = FALSE;      /* save HMM defs in binary */
static Boolean saveGlobOpts = TRUE;     /* save ~o with HMM defs */
static Boolean saveRegTree = FALSE;    /* save regression classes and tree */ 
static Boolean saveBaseClass = FALSE;   /* save base classes */ 
static Boolean saveInputXForm = FALSE;  /* save input xforms with models set */
static Boolean forceHSKind= FALSE;      /* force HMM Set Kind */
static Boolean keepDistinct=FALSE;      /* keep orphan HMMs distinct */
static Boolean discreteLZero=FALSE;     /* map DLOGZERO to LZERO */
static Boolean reorderComps=FALSE;      /* re-order mixture components (PDE) */

static Boolean allowOthers=TRUE;        /* allow unseen models in files */
static HSetKind cfHSKind;
static char orphanMacFile[100];         /* last resort file for new macros */

static XFDirLink xformDirNames = NULL;  /* linked list of input transform directories */
static Boolean indexSet = FALSE;        /* have the indexes been set for the model set */

static MemHeap xformStack;              /* For Storage of xforms with no model sets ... */

static int pde1BlockEnd = 13;          /* size of PDE blocks */
static int pde2BlockEnd = 26;          /* size of PDE blocks */
static LogFloat pdeTh1 = -5.0;         /* threshold for 1/3 PDE */
static LogFloat pdeTh2 = 0.0;          /* threshold for 2/3 PDE */

#ifdef PDE_STATS
static int nGaussTot = 0;
static int nGaussPDE1 = 0;
static int nGaussPDE2 = 0;
#endif

void InitSymNames(void);

/* EXPORT->InitModel: initialise memory and configuration parameters */
void InitModel(void)
{
   int i;
   double d;
   Boolean b;
   char buf[MAXSTRLEN];
   
   Register(hmodel_version,hmodel_vc_id);
   CreateHeap(&xformStack,"XFormStore",MSTAK, 1, 0.5, 100 ,  1000 );
   strcpy(orphanMacFile,"newMacros");
   InitSymNames();
   nParm = GetConfig("HMODEL", TRUE, cParm, MAXGLOBS);
   if (nParm>0){
      if (GetConfInt(cParm,nParm,"TRACE",&i)) trace = i;
      if (GetConfBool(cParm,nParm,"CHKHMMDEFS",&b)) checking = b;
      if (GetConfBool(cParm,nParm,"SAVEBINARY",&b)) saveBinary = b;
      if (GetConfBool(cParm,nParm,"KEEPDISTINCT",&b)) keepDistinct = b;
      if (GetConfBool(cParm,nParm,"SAVEGLOBOPTS",&b)) saveGlobOpts = b;
      if (GetConfBool(cParm,nParm,"SAVEREGTREE",&b)) saveRegTree = b;
      if (GetConfBool(cParm,nParm,"SAVEBASECLASS",&b)) saveBaseClass = b;
      if (GetConfBool(cParm,nParm,"SAVEINPUTXFORM",&b)) saveInputXForm = b;
      if (GetConfBool(cParm,nParm,"ALLOWOTHERHMMS",&b)) allowOthers = b;
      if (GetConfBool(cParm,nParm,"DISCRETELZERO",&b)) discreteLZero = b;
      if (GetConfStr (cParm,nParm,"ORPHANMACFILE",buf))
         strcpy(orphanMacFile,buf);
      if (GetConfStr (cParm,nParm,"HMMSETKIND",buf)) {
         if (strcmp(buf,"PLAIN")==0) cfHSKind = PLAINHS;
         else if (strcmp(buf,"SHARED")==0) cfHSKind = SHAREDHS;
         else if (strcmp(buf,"TIED")==0) cfHSKind = TIEDHS;
         else if (strcmp(buf,"DISCRETE")==0) cfHSKind = DISCRETEHS;
         else
            HError(7070,"InitModel: Unknown HMM kind %s",buf);
         forceHSKind = TRUE;
      }
      if (GetConfBool(cParm,nParm,"REORDERCOMPS",&b)) reorderComps = b;
      if (GetConfInt(cParm,nParm,"PDE1BLOCKEND",&i)) pde1BlockEnd = i;
      if (GetConfInt(cParm,nParm,"PDE2BLOCKEND",&i)) pde2BlockEnd = i;
      if (GetConfFlt(cParm,nParm,"PDETHRESHOLD1",&d)) pdeTh1 = d;
      if (GetConfFlt(cParm,nParm,"PDETHRESHOLD2",&d)) pdeTh2 = d;
   }
}

/* -------------------- Check Model Consistency -------------------- */

/* CheckMix: check given mixture pdf of state n, stream s, mixture m */
static ReturnStatus CheckMix(char *defName, MixPDF *mp, int n, int s, int m, int sw)
{  
   if (trace&T_CHK) {
      printf("HModel:       checking mix %d\n",m);  fflush(stdout);
   }
   if (mp->mean == NULL){
      HRError(7030,"CheckMix: %s: pdf for s=%d,j=%d,m=%d has NULL mean",
              defName,n,s,m);
      return(FAIL);
   }
   if (VectorSize(mp->mean) != sw){
      HRError(7030,"CheckMix: %s: mean for s=%d,j=%d,m=%d has bad vecSize",
              defName,n,s,m);
      return(FAIL);
   }
   if (mp->cov.var == NULL){
      HRError(7030,"CheckMix: %s: pdf for s=%d,j=%d,m=%d has NULL covariance",
              defName,n,s,m);
      return(FAIL);
   }
   
   switch(mp->ckind){
   case DIAGC:
      if (VectorSize(mp->cov.var) != sw){
         HRError(7030,"CheckMix: %s: var for s=%d,j=%d,m=%d has bad vecSize",
                 defName,n,s,m);
         return(FAIL);
      }
      break;
   case FULLC:
   case LLTC:
      if (TriMatSize(mp->cov.inv) != sw){
         HRError(7030,"CheckMix: %s: inv for s=%d,j=%d,m=%d has bad dimens",
                 defName,n,s,m);
         return(FAIL);
      }
      break;
   case XFORMC:
      if (NumCols(mp->cov.xform) != sw){
         HRError(7030,"CheckMix: %s: xform for s=%d,j=%d,m=%d has bad dimens",
                 defName,n,s,m);
         return(FAIL);
      }
      break;
   }
   if (mp->gConst == LZERO){
      if (mp->ckind==DIAGC)
         FixDiagGConst(mp);
      else if (mp->ckind==FULLC)
         FixFullGConst(mp,-CovDet(mp->cov.inv));
      else if (mp->ckind==LLTC)
         FixLLTGConst(mp);
      else if (mp->ckind==XFORMC)
         HRError(7030,"CheckMix: No GConst set for xformc covariance matrix");
   }
   Touch(&mp->nUse);
   return(SUCCESS);
}

/* CheckStream: check the stream s for state n */
static ReturnStatus CheckStream(char *defName, HLink hmm, StreamElem *se, int s, int n)
{
   double sum=0.0;
   int m,sw;
   LogFloat wt;
   MixtureElem *me;
   MixPDF *mp;
   HSetKind  hk;
   
   if (trace&T_CHK) {
      printf("HModel:    checking stream %d, se=%p\n",s,se); 
      fflush(stdout);
   }
   hk = hmm->owner->hsKind;
   sw = hmm->owner->swidth[s];
   switch(hk){
   case PLAINHS:
   case SHAREDHS:
      me = se->spdf.cpdf+1;
      for (m=1; m<=se->nMix; m++,me++){
         wt=me->weight; wt = MixWeight(hmm->owner,wt);
         sum += wt;
         if (wt > MINMIX ){
            if (me->mpdf == NULL){
               HRError(7030,"CheckStream: %s: MixDef %d missing for s=%d, j=%d",
                       defName, m, s, n);
               return(FAIL);
            }
            mp = me->mpdf;
            if((!IsSeen(mp->nUse))&&(CheckMix(defName,me->mpdf,n,s,m,sw)<SUCCESS))
               return(FAIL);
         }
      }
      break;
   case TIEDHS:
      for (m=1; m<=se->nMix; m++)
         sum += se->spdf.tpdf[m];
      break;
   case DISCRETEHS:
      for (m=1; m<=se->nMix; m++)
         sum += exp((float)se->spdf.dpdf[m]/DLOGSCALE);
      break;
   }
   if (sum<0.99 || sum>1.01){
      HRError(7031,"CheckStream: %s: Mix weights sum %e for s=%d, j=%d",
              defName,sum,s,n);
      return(FAIL);
   }
   return(SUCCESS);
}

/* CheckState: check state n */
static ReturnStatus CheckState(char *defName, HLink hmm, StateInfo *si, int n)
{
   int s,S;
   StreamElem *ste;
   
   if (trace&T_CHK) {
      printf("HModel:  checking state %d, si=%p\n",n,si); fflush(stdout);
   }
   S = hmm->owner->swidth[0];
   if (si->pdf == NULL){
      HRError(7030,"CheckState: %s: state %d has no pdf",defName,n);
      return(FAIL);
   }
   ste = si->pdf+1;
   for (s=1; s<=S; s++,ste++){
      if (ste->spdf.cpdf == NULL){
         HRError(7030,"CheckState: %s: Stream %d missing in state %d",defName,s,n);
         return(FAIL);
      }
      if(CheckStream(defName,hmm,ste,s,n)<SUCCESS)
         return(FAIL);
   }
   if (S>1)
      if (si->weights==NULL){
         HRError(7030,"CheckState: %s: Stream weights missing in state %d",defName,n);
         return(FAIL);
      }
   Touch(&si->nUse);
   return(SUCCESS);
}

/* CheckHMM: check the consistency of the given model */
static ReturnStatus CheckHMM(char *defName, HLink hmm)
{
   int i;
   StateElem *se;
   StateInfo *si;
   
   if (trace&T_CHK) {
      printf("HModel: checking HMM %s\n",defName); fflush(stdout);
   }
   for (i=2,se=hmm->svec+2; i<hmm->numStates; i++,se++) {
      si = se->info;
      if (si == NULL){
         HRError(7030,"CheckHMM: %s: state %d has NULL info",defName,i);
         return(FAIL);
      }
      else if((!IsSeen(si->nUse))&&(CheckState(defName,hmm,si,i)<SUCCESS))
         return(FAIL);
   }
   return(SUCCESS);
}

/* CheckTMRecs: check the tied mix codebook attached to a HMM Set */
static ReturnStatus CheckTMRecs(HMMSet *hset)
{
   int s,m,sw;
   MixPDF *mp;
   
   if (trace&T_CHK) {
      printf("HModel: checking tied mixture codebook\n"); fflush(stdout);
   }
   for (s=1;s<=hset->swidth[0]; s++){
      sw = hset->swidth[s];
      if (hset->tmRecs[s].mixId == NULL){
         HRError(7030,"CheckTMRecs: no mix id set in stream %d",s);
         return(FAIL);
      }
      if (hset->tmRecs[s].mixes == NULL){
         HRError(7030,"CheckTMRecs: no mixes array allocated in stream %d",s);
         return(FAIL);
      }
      for (m=1; m<=hset->tmRecs[s].nMix; m++){
         mp = hset->tmRecs[s].mixes[m];
         if(CheckMix("TMRec",mp,0,s,m,sw)<SUCCESS)
            return(FAIL);
      }
   }
   return(SUCCESS);
}

/* CheckDiscrete: check discrete HMM Set */
static ReturnStatus CheckDiscrete(HMMSet *hset)
{
   int s;

   for (s=1;s<=hset->swidth[0]; s++){
      if ((hset->swidth[s] != 1) && (!(HasVQ(hset->pkind)))) {
         HRError(7030,"CheckDiscrete: stream width not equal to 1 in discrete stream %d ",s);
         return(FAIL);
      }
   }
   return (SUCCESS);
}

/* CheckHSet: check the consistency of a complete HMM Set */
static ReturnStatus CheckHSet(HMMSet *hset)
{
   int h;
   MLink m;
   
   for (h=0; h<MACHASHSIZE; h++)
      for (m=hset->mtab[h]; m!=NULL; m=m->next)
         if (m->type == 'h')
            if(CheckHMM(m->id->name,(HLink)m->structure)<SUCCESS)
               return(FAIL);
   if ((hset->hsKind == TIEDHS) && (CheckTMRecs(hset)<SUCCESS))
      return(FAIL);
   if ((hset->hsKind == DISCRETEHS) && (CheckDiscrete(hset)<SUCCESS))
      return(FAIL);

   ClearSeenFlags(hset,CLR_ALL);
   return(SUCCESS);

}

/* ------------------------ Lexical Scanner ------------------------ */

#define MAXSYMLEN 40

/* Internal and binary keyword representation */
typedef enum {   /* Only a character big !! */
   BEGINHMM, USEMAC, ENDHMM, NUMMIXES, 
   NUMSTATES, STREAMINFO, VECSIZE, 
   NDUR, PDUR, GDUR, RELDUR, GENDUR,
   DIAGCOV,  FULLCOV, XFORMCOV,
   STATE, TMIX, MIXTURE, STREAM, SWEIGHTS,
   MEAN, VARIANCE, INVCOVAR, XFORM, GCONST,
   DURATION, INVDIAGCOV, TRANSP, DPROB, LLTCOV, LLTCOVAR,
   XFORMKIND=90, PARENTXFORM, NUMXFORMS, XFORMSET,
   LINXFORM, OFFSET, BIAS, LOGDET, BLOCKINFO, BLOCK, BASECLASS, 
   CLASS, XFORMWGTSET, CLASSXFORM, MMFIDMASK, PARAMETERS,
   NUMCLASSES, ADAPTKIND, PREQUAL, INPUTXFORM,
   RCLASS=110, REGTREE, NODE, TNODE,
   HMMSETID=119,
   PARMKIND=120, 
   MACRO, EOFSYM, NULLSYM   /* Special Syms - not literals */
} Symbol;

/* Mapping between verbose keywords for readable external HMM format */
/*  and internal/external binary token numbers */
static struct {
   char *name;
   Symbol sym; } symMap[] = 
   { { "BEGINHMM", BEGINHMM }, { "USE", USEMAC }, 
     { "ENDHMM", ENDHMM }, { "NUMMIXES", NUMMIXES }, 
     { "NUMSTATES", NUMSTATES }, { "STREAMINFO", STREAMINFO }, 
     { "VECSIZE", VECSIZE }, { "NULLD", NDUR }, 
     { "POISSOND", PDUR }, { "GAMMAD", GDUR },