/* ----------------------------------------------------------- *//*                                                             *//*                          ___                                *//*                       |_| | |_/   SPEECH                    *//*                       | | | | \   RECOGNITION               *//*                       =========   SOFTWARE                  */ /*                                                             *//*                                                             *//* ----------------------------------------------------------- *//*         Copyright: Microsoft Corporation                    *//*          1995-2000 Redmond, Washington USA                  *//*                    http://www.microsoft.com                 *//*                                                             *//*   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: HAdapt.h      Adaptation Library module       *//* ----------------------------------------------------------- *//* !HVER!HAdapt:   3.2 [CUED 09/12/02] */#ifndef _HADAPT_H_#define _HADAPT_H_#ifdef __cplusplusextern "C" {#endifenum _RegClassType{ADPTUNDEF=0,ADPTFIXED=1,ADPTTREE=2};typedef enum _RegClassType RegClassType;enum _RegTransType{TRANSUNDEF=0,MEANONLY=1,MEANVAR=2};typedef enum _RegTransType RegTransType;#define DEF_BLOCKSIZE          1   /* define the default blocksize */#define DEF_REGCLASS    ADPTTREE   /* define the default Reg Class Type *//* ------------- Block Matrix Definitions -------------- *//* A block matrix is a array of square matrices *//* There are nBlocks matrices,    and each matrix is a square matrix of blockSize *//* [1..nBlocks] of matrix[blockSize][blockSize] */typedef Matrix  *BlockMatrix;typedef DMatrix *BlockDMatrix;typedef TriMat  *BlockTriMatrix;/* ------------ End Block Matrix Definitions ------------ *//* structure to hold the mean regression matrix & bias vector: x' = ax + b   plus the Z accumumulation structure */typedef struct {  BlockMatrix a;            /* block diagonal matrix */  Vector b;                 /* bias vector           */} OffsetBMat;/* specialist structure for the G matrix ! */typedef struct {  BlockTriMatrix a;         /* block tri-diagonal matrix */  Vector b;                 /* bias vector               */} OffsetTriBMat;typedef struct {  Boolean speechFlag;       /* flag to see if class is speech/non-speech */  short baseformIndex;      /* regression class index number */  int nComponents;          /* number of components in this class */  float  occ;               /* baseform's occ count for ALL incoming data */    Vector H;                 /* the H accumulates for base class (covar) */  OffsetTriBMat **G;        /* the G accumulates for base class */  OffsetBMat *Z;            /* the Z accumulates for base class */  MixPDF **pdfList;         /* array list of pointers to pdfs in base class */} BaseformAcc;typedef struct {  short        nBases;      /* number of baseforms for this node */  short     nodeIndex;      /* node's index number */  short        *bases;      /* list of baseform indices for this node */  int       nodeComps;      /* the number of components present at a node */    float       nodeOcc;      /* node's occupation count */      OffsetBMat  *WTrans;      /* mean matrix transform for node */   Vector       HTrans;      /* diag cov vector transform for node */  Matrix    backTrans;      /* inverse mean transform matrix */} RegNode ;/* structure to hold the regression tree :-   a) Global transform has all stats stored at the root node    b) Fixed transform has only left children (all right children are NULL)   c) Standard regression tree */typedef struct _RegTree {    struct _RegTree *left;  struct _RegTree *right;  RegNode *nodeInfo;  Ptr node;} RegTree;/* structure to hold information pertaining to this transform set */typedef struct {  char *uid;                /* user identifier */   char *name;               /* user's full name */  char *hmmSetId;           /* hmm set identifier */  char *rcId;               /* regression class type identifier */  char *chan;               /* desc of channel used for data collection */  char *desc;               /* general description */} TransformId;/* structure to hold the regression classes and accumulation information */typedef struct {  TransformId *transId;     /* information identifier for this transform */  HMMSet *hset;             /* HMMSet being transformed */  MemHeap *hmem;            /* mem heap for HAdapt module */  MemHeap pdfStatsMem;      /* mem heap solely for the mixPDF accumulations */  MemHeap transMem;         /* mem heap solely for the transform structures */  short vSize;              /* vector size */  short nBlocks;            /* number of blocks in each transform */   short nBaseTransforms;    /* total number of possible transforms   */  RegClassType classKind;   /* type of regression class */  RegTransType transKind;   /* type of regression transformations */  TriState adptSil;         /* Flag for adapting silence */  float nodeOccThresh;      /* Node occupation threshold for 			       generating a transform */  /* Boolean transWithBias; Use of loaded bias */  /*  TriState useBias;     Use of bias */  RegTree *rtree;           /* regression tree */  BaseformAcc **baseforms;  /* list of baseform accumulates */} RegTransInfo ;/* structure to store the regression accumulations */typedef struct {  /* in place of PreComp -- necessary for f-b routines */  /* DO NOT MOVE or REMOVE !!! */  PreComp  pre;             /* PreComputed Mixture calculation structure */  /* the rest */  float    occ;             /* occupation count */  Vector   obsSum;          /* summed speech vectors, scaled by posterior */  Vector   obsSqSum;        /* summed sqared sp. vecs, scaled by posterior */  /* MixPDF   *nextMix;      next mixture in the regression class list */} RegAcc ;/* -------------------- Initialisation Functions -------------------------- */void InitAdapt(void);/*   Initialise configuration parameters*/void InitialiseTransform( HMMSet *hset, MemHeap *x, RegTransInfo *rt,			  Boolean adapt );/*   Initialise transforms storage and grouping of hmmset components   to regression base class using linked lists */void InitialiseAdapt(HMMSet *hset,MemHeap *x, RegTransInfo *rt);/*    Initialise adaptation storage*//* -------------------- Save/Load Transform Functions --------------------- */void SaveTransformSet(HMMSet *hset, RegTransInfo *rt, char *saveFile, 		      char *transPath, char *uid, char *uname, char *chan, 		      char *desc, Boolean saveStats, Boolean global,		      Boolean saveBinary);/*    Save the transforms to a file, and also optionally save   transform statistics allowing continued adaptation*/Boolean LoadTransformSet(HMMSet *hset, char *tfile, char *uid,			 RegTransInfo *rt, Boolean *useStats);/*    Load the transforms into memory; if useStats is requested and    transform does not contain any statistics, then useStats is   reset to FALSE -- returns FALSE if tmf is unsuitable   for current HMMSet*//* ---------------- Accumulation Control Functions ------------------------ */void AccAdaptFrame(double Lr, Vector speechVec, MixPDF *mp, RegTransInfo *rt);/*    Accumulate frame stats into specific mixture comp*/void ClearRegCompStats(HMMSet *hset, RegTransInfo *rt);/*    Clear for regression level accumulated stats */void ClearBaseClassStats(RegTransInfo *rt);/*    Clear for base class level accumulated stats *//* ----------------- Mixture Transformation Functions --------------------- *//* ----- Transform application functions ---- */void ApplyMeanTransforms(RegTransInfo *rt, RegTree *t);/*    Apply the mean transform to the regression classes*/void ApplyCovarTransforms(RegTransInfo *rt, RegTree *t);/*   Apply the covariance transform to the regression classes*/Boolean ApplyMeanGlobalTransform(RegTransInfo *rt);/*    Apply the mean Global transformation -- returns true if global is found */Boolean ApplyCovarGlobalTransform(RegTransInfo *rt);/*   Apply the covar Global transformation -- returns true if global is found */void ApplyTransforms(RegTransInfo *rt);/*    Apply the mean and possibly the covar MLLR transformations*/Boolean ApplyBackwardGlobalTransform(RegTransInfo *rt);/*    Apply the backward global transformation to the model set    to obtain the "original" models -- returns TRUE if global is found*/void ApplyBackwardTransforms(RegTransInfo *rt);/*   Apply the backward transformation to the model set    to obtain the "original" models*//* ----- Transform calculation/update ----- */void CalcMeanTransforms(RegTransInfo *rt);/*    Calculate the mean transforms for every regression class */void CalcCovarTransforms(RegTransInfo *rt);/*    Calculate the covariance transforms for every regression class*/void GetBackwardTransforms(RegTransInfo *rt);/*    Calculates the backward transformation by inverting    the original transform*/void UpdateMAP(RegTransInfo *rt, float tau);/*    Update the models using Maximum A-Posteriori Adaptation (MAP) *//* ---------------- General adaptation Call ------------------------------- *//* EXPORT->DoAdaptation: Given the initialisation to the adaptation *//* MLLR Adaptation    Before this function is called HFB must be called via "FBFile"   to perfrom a frame/state alignment and accumulate regression statistics   at the mixture level    DoAdaptation does:-   1) Transform current models back to their original state (as loaded)   2) Calculation mean transform and apply the transform   3) Calculate the covariance transform (if requested) and apply*/void DoAdaptation(RegTransInfo *rt, Boolean global);#ifdef __cplusplus}#endif#endif  /* _HADAPT_H_ *//* ---------------------------- END HAdapt.h ------------------------------ */