语音识别之HTK入门（四）——HCompV源码解析

上一篇博客中用HCompV工具通过全局数据来初始化HMM模型参数，包括proto和vFloor两个文件。同时手工制作hmmdefs模型，在这个文件中列举了每一个单因素模型的参数，它们是完全相同的。

现在一步一步看看HCompV是如何达成目标的。

如前面博客指出的，可以通过visual studio跟踪调试HCompV。

首先列出大体的流程涉及的函数：

分别是：

1）InitShell    // 对命令行参数进行处理，这个函数在多个HTKTools之间都用到，它主要把参数保存到savedCommandLine这里，后续对它进行读写处理；比如系统的参数设置，一般是“-C|S|V|D”，并把余下的参数保存到arglist二维字符数组中，计数变量为argcount；

2）然后是一堆初始化工作

InitMem();   
InitLabel();
InitMath();  
InitSigP();
InitWave();  
InitAudio();
InitVQ();    
InitModel();
if(InitParm()<SUCCESS)  
    HError(2000,"HCompV: InitParm failed");
if (!InfoPrinted() && NumArgs() == 0)
    ReportUsage();
if (NumArgs() == 0) Exit(0);

3）核心的处理函数是从CreateHMMSet开始的。

下面来详细分析下这个函数做了哪些任务，首先在代码中函数声明为

/* EXPORT->CreateHMMSet: create the basic HMMSet structure */
void CreateHMMSet(HMMSet *hset, MemHeap *heap, Boolean allowTMods)；它接受三个参数分别是HMMSet结构指针、MemHeap结构指针和allowTMods的布尔值。因此这两个结构大体长什么样子肯定要有所了解，否则后面的操作、赋值、计算都不明了。而allowTMods是个开关量。

typedef struct _HMMSet{
   MemHeap *hmem;          /* memory heap for this HMM Set */   
   Boolean *firstElem;     /* first element added to hmem during MakeHMMSet*/
   char *hmmSetId;         /* identifier for the hmm set */
   MILink mmfNames;        /* List of external file names */
   int numLogHMM;          /* Num of logical HMM's */
   int numPhyHMM;          /* Num of distinct physical HMM's */
   int numFiles;           /* total number of ext files */
   int numMacros;          /* num macros used in this set */
   MLink * mtab;           /* Array[0..MACHASHSIZE-1]OF MLink */
   PtrMap ** pmap;         /* Array[0..PTRHASHSIZE-1]OF PtrMap* */
   Boolean allowTMods;     /* true if HMMs can have Tee Models */
   Boolean optSet;         /* true if global options have been set */
   short vecSize;          /* dimension of observation vectors */
   short swidth[SMAX];     /* [0]=num streams,[i]=width of stream i */
   ParmKind pkind;         /* kind of obs vector components */
   DurKind dkind;          /* kind of duration model (model or state) */
   CovKind ckind;          /* cov kind - only global in V1.X */
   HSetKind hsKind;        /* kind of HMM set */
   TMixRec tmRecs[SMAX];   /* array[1..S]of tied mixture record */
   int numStates;          /* Number of states in HMMSet */
   int numSharedStates;    /* Number of shared states in HMMSet */
   int numMix;             /* Number of mixture components in HMMSet */
   int numSharedMix;       /* Number of shared mixtures in HMMSet */
   int numTransP;          /* Number of distinct transition matrices */
   int ckUsage[NUMCKIND];  /* Number of components using given ckind */
   InputXForm *xf;         /* Input transform of HMMSet */
   AdaptXForm *semiTied;   /* SemiTied transform associated with model set */
   short projSize;         /* dimension of vector to update */

   /* Adaptation information accumulates */
   Boolean attRegAccs;   /* have the set of accumulates been attached */
   Boolean attXFormInfo; /* have the set of adapt info been attached */
   Boolean attMInfo;     /* have the set of adapt info been attached */
   AdaptXForm *curXForm;
   AdaptXForm *parentXForm;
   
   /* Added to support LogWgts */
   Boolean logWt;       /* Component weights are stored as Logs */

   /* Added to support delayed loading of the semi-tied transform */
   char *semiTiedMacro;  /* macroname of semi-tied transform */

} HMMSet;

看每个数据项的注释，大体意思应该比较明白了。其中第一项就是HHMSet需要涉及的内存空间，用来保存模型参数，也就是这个函数第二个形参的目的。

typedef struct {
   char *name;          /*            name of this memory heap            */
   HeapType type;       /*              type of this heap                 */
   float growf;         /*           succ blocks grow as 1+growf          */
   size_t elemSize;     /*  size of each elem              1 always       */
   size_t minElem;      /*  init #elems per blk      init #bytes per blk  */
   size_t maxElem;      /*  max #elems per block     max #bytes per blk   */
   size_t curElem;      /*  current #elems per blk   curr #bytes per blk  */
   size_t totUsed;      /*  total #elems used        total #bytes used    */
   size_t totAlloc;     /*  total #elems alloc'ed    total #bytes alloc'd */
   BlockP heap;         /*               linked list of blocks            */
   Boolean protectStk;  /*  MSTAK only, prevents disposal below Stack Top */
}MemHeap;

该HMMSet定义了hmm模型的个数，状态数，观察向量的维度转移概率矩阵的个数等等。在这个函数里只是初始化这样的结构体，而具体值需要后面根据模型文件来构建模型时填充。

 CreateHMMSet(&hset,&gstack,FALSE);
   pathPattern[0]='\0';
   while (NextArg() == SWITCHARG) {
      s = GetSwtArg();
      if (strlen(s)!=1) 
         HError(2019,"HCompV: Bad switch %s; must be single letter",s);
      switch(s[0]){
      case 'f':
         if (NextArg() != FLOATARG)
            HError(2019,"HCompV: Variance floor scale expected");
         vFloorScale = GetChkedFlt(0.0,100.0,s);
         break;
      case 'l':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: Segment label expected");
         segLab = GetStrArg();
         break;
      case 'm':
         meanUpdate = TRUE;
         break;
      case 'o':
         outfn = GetStrArg();
         break;     
      case 'v':
         if (NextArg() != FLOATARG)
            HError(2019,"HCompV: Minimum variance level expected");
         minVar = GetChkedFlt(0.0,100.0,s);
         break;
      case 'k':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: speaker pattern expected");
         strcpy(spPattern,GetStrArg());
         if (strchr(spPattern,'%')==NULL)
            HError(2019,"HCompV: Speaker mask invalid");
         break;
      case 'c':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: CMV output dir expected");
         strcpy(cmDir,GetStrArg());
         DoCMV = TRUE;
         break;
      case 'p':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: path pattern expected");
         strcpy(pathPattern,GetStrArg());
         if (strchr(pathPattern,'%')==NULL)
            HError(2019,"HCompV: Path mask invalid");
         break;
      case 'q':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: output flags (nmv)");
         strcpy(oflags,GetStrArg());
         break;
      case 'B':
         saveBinary = TRUE;
         break;
      case 'F':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: Data File format expected");
         if((dff = Str2Format(GetStrArg())) == ALIEN)
            HError(-2089,"HCompV: Warning ALIEN Data file format set");
         break;
      case 'G':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: Label File format expected");
         if((lff = Str2Format(GetStrArg())) == ALIEN)
            HError(-2089,"HCompV: Warning ALIEN Label file format set");
         break;
      case 'H':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: HMM macro file name expected");
         AddMMF(&hset,GetStrArg());
         break;
      case 'I':
         if (NextArg() != STRINGARG)
            HError(2019,"HCompV: MLF file name expected");
         LoadMasterFile(GetStrArg());
         break;
      case 'L':
         if (NextArg()!=STRINGARG)
            HError(2019,"HCompV: Label file directory expected");
         labDir = GetStrArg();
         break;
      case 'M':
         if (NextArg()!=STRINGARG)
            HError(2019,"HCompV: Output macro file directory expected");
         outDir = GetStrArg();
         break;
      case 'T':
         if (NextArg() != INTARG)
            HError(2019,"HCompV: Trace value expected");
         trace = GetChkedInt(0,077,s); 
         break;
      case 'X':
         if (NextArg()!=STRINGARG)
            HError(2019,"HCompV: Label file extension expected");
         labExt = GetStrArg();
         break;
      default:
         HError(2019,"HCompV: Unknown switch %s",s);
      }
   }

   /* if not doing CMV, do standard HCompV */
   if (DoCMV == FALSE){
      if (NextArg()!=STRINGARG)
         HError(2019,"HCompV: Source HMM file name expected");
      hmmfn = GetStrArg();
      Initialise();
      do {
         if (NextArg()!=STRINGARG)
            HError(2019,"HCompV: Training data file name expected");
         datafn = GetStrArg();
         LoadFile(datafn);
      } while (NumArgs()>0);
      SetCovs();
      FixGConsts(hmmLink);
      SaveModel(outfn);   
      if (trace&T_TOP)
         printf("Output written to directory %s\n",(outDir==NULL)?"./":outDir);
      if (vFloorScale>0.0)
         PutVFloor();
   }

while训练语句就是在处理命令行的参数，不断的读取由initshell处理后剩下的参数，在HCompV命令中，有-f 0.01 -m -M dir proto等7个参数，处理完毕后，设置了全局变量vFloorScale（协方差的最低值系数），meanUpdate（bool，同时更新期望值），outDir（模型的输出目录）和hmmfn（描述模型的文件名）。

接着，默认情况下DoCMV为false，因此进入if第一个分支，设置hmmfn值为proto，它就是我们之前为HMM模型描述的结构。然后执行Initialise()函数。

4）Initialise()

函数前面的注释 /* Initialise: load HMMs and create accumulators */

下面把这个函数的主要调用关系罗列出来，然后分析它们各自都干了什么。

void Initialise(void)
{
   int s,V;
   Boolean eSep;
   char base[MAXSTRLEN];
   char path[MAXSTRLEN];
   char ext[MAXSTRLEN];

   /* Load HMM defs */     
   if(MakeOneHMM(&hset,BaseOf(hmmfn,base))<SUCCESS)
      HError(2028,"Initialise: MakeOneHMM failed");
   if(LoadHMMSet(&hset,PathOf(hmmfn,path),ExtnOf(hmmfn,ext))<SUCCESS)
      HError(2028,"Initialise: LoadHMMSet failed");
   SetParmHMMSet(&hset);

   /* Create a heap to store the input data */
   CreateHeap(&iStack,"InBuf", MSTAK, 1, 0.5, 100000, LONG_MAX);
   
   /* Get a pointer to the physical HMM */
   hmmId = GetLabId(base,FALSE);
   macroLink = FindMacroName(&hset,'h',hmmId);
   hmmLink = (HLink)macroLink->structure;

   /* Find out for which streams full covariance is needed */
   CheckVarianceKind( );

   /* Create accumulators for the mean and variance */
   for (s=1;s<=hset.swidth[0]; s++){
      V = hset.swidth[s];
      accs[s].meanSum=CreateVector(&gstack,V);
      ZeroVector(accs[s].meanSum);
      if (fullcNeeded[s]) {
         accs[s].squareSum.inv=CreateSTriMat(&gstack,V);
         accs[s].fixed.inv=CreateSTriMat(&gstack,V);
         ZeroTriMat(accs[s].squareSum.inv);
      }
      else {
         accs[s].squareSum.var=CreateSVector(&gstack,V);
         accs[s].fixed.var=CreateSVector(&gstack,V);
         ZeroVector(accs[s].squareSum.var);
      }
   }

   /* Create an object to hold the input parameters */
   SetStreamWidths(hset.pkind,hset.vecSize,hset.swidth,&eSep);
   obs=MakeObservation(&gstack,hset.swidth,hset.pkind,FALSE,eSep);
   if(segLab != NULL) {
      segId = GetLabId(segLab,TRUE);
   }

}

可以看出首先调用MakeOneHMM，看看它完成了哪些操作。

/* EXPORT->MakeOneHMM: Create a singleton for the HMM hname */
ReturnStatus MakeOneHMM(HMMSet *hset, char *hname)
{
   if(InitHMMSet(hset, hname, TRUE)<SUCCESS){
      ResetHMMSet(hset);
      return(FAIL);
   }
   return(SUCCESS);
}

它实际上调用了初始化InitHMMSet函数，通过读取proto这个hmm模型文件。proto文件的内容如下：

~o <VecSize> 39 <MFCC_0_D_A>
~h "proto"
<BeginHMM>
  <NumStates> 5
  <State> 2
    <Mean> 39
      0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
    <Variance> 39
      1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
 <State> 3
    <Mean> 39
      0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
    <Variance> 39
      1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
 <State> 4
    <Mean> 39
      0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 
    <Variance> 39
      1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
 <TransP> 5
  0.0 1.0 0.0 0.0 0.0
  0.0 0.6 0.4 0.0 0.0
  0.0 0.0 0.6 0.4 0.0
  0.0 0.0 0.0 0.7 0.3
  0.0 0.0 0.0 0.0 0.0
<EndHMM>

其实这个hmm原型文件描述了HMM主要的参数，包括5个状态，观察向量的维度是39，它是语音的MFCC以及一阶、二阶参数（MFCC_0_D_A）。这个原型文件只描述了一个hmm模型，名字就是"proto"。

在InitHMMSet函数中调用了CreateHMM函数完成模型的单个模型的构建。模型名称为“proto”，并保存到HMMSet中，但是实际上这时这个模型的参数没有设定，只是有了一个名字的空模型。

在接下来的循环语句中，LoadFile函数读取每个MFCC文件，并计算特征向量的均值和方差，通过累加的方式。

SetCovs()函数完成方差的计算，并复制给每个状态，来初始化状态的概率密度函数。

/* SetCovs: set covariance values in hmm */
void SetCovs(void)
{
   int i,s,m;
   StateElem *se;
   StreamElem *ste;
   MixtureElem *me;
   MixPDF *mp;

   CalcCovs();  // 计算均值和方差 保存在全局的CovAcc对象中。
   if (trace&T_TOP) {
      printf("Updating HMM ");
      if (meanUpdate) printf("Means and ");
      printf("Covariances\n");
   }
   for (i=2,se=hmmLink->svec+2; i < hmmLink->numStates; i++,se++)
      for (s=1,ste=se->info->pdf+1; s <= hset.swidth[0]; s++,ste++)
         for (m=1,me = ste->spdf.cpdf+1; m<=ste->nMix; m++, me++) {
            mp = me->mpdf;
            if (meanUpdate && !IsSeenV(mp->mean)){     /* meanSum now holds mean */
               CopyVector(accs[s].meanSum,mp->mean);   // 把计算的均方差复制给模型的状态
               TouchV(mp->mean);
            }
            if (!IsSeenV(mp->cov.var)){
               if (mp->ckind==FULLC)
                  CopyMatrix(accs[s].fixed.inv,mp->cov.inv);
               else if (fullcNeeded[s])  /* dont need full cov, but its all we have */                
                  TriDiag2Vector(accs[s].fixed.inv,mp->cov.var);
               else
                  CopyVector(accs[s].fixed.var,mp->cov.var);
               TouchV(mp->cov.var);
            }
         }
   ClearSeenFlags(&hset,CLR_ALL);
}