read_binhmm.c

julius version 4.12.about sound recognition.

/**
 * @file   read_binhmm.c
 * 
 * <JA>
 * @brief  バイナリ妨及の %HMM 年盗ファイルを粕み哈む
 *
 * Julius は迫极のバイナリ妨及の %HMM 年盗ファイルをサポ〖トしていますˉ
 * HTKのアスキ〖妨及の %HMM 年盗ファイルからバイナリ妨及への恃垂は·
 * 身掳のツ〖ル mkbinhmm で乖ないますˉこのバイナリ妨及は·HTK の
 * バイナリ妨及とは润高垂ですので庙罢して布さいˉ
 * </JA>
 * 
 * <EN>
 * @brief  Read a binary %HMM definition file
 *
 * Julius supports a binary format of %HMM definition file.
 * The tool "mkbinhmm" can convert the ascii format HTK %HMM definition
 * file to this format.  Please note that this binary format is 
 * not compatible with the HTK binary format.
 * </EN>
 * 
 * @author Akinobu LEE
 * @date   Wed Feb 16 05:23:59 2005
 *
 * $Revision: 1.5 $
 * 
 */
/*
 * Copyright (c) 2003-2005 Shikano Lab., Nara Institute of Science and Technology
 * Copyright (c) 2005-2007 Julius project team, Nagoya Institute of Technology
 * All rights reserved
 */

#include <sent/stddefs.h>
#include <sent/htk_param.h>
#include <sent/htk_hmm.h>

#undef DMES			/* define to enable debug message */

static boolean gzfile;	      ///< TRUE when opened by fopen_readfile

#define rdn(A,B,C,D) if (rdnfunc(A,B,C,D) == FALSE) return FALSE
#define rdn_str(A,B,C) if ((C = rdn_strfunc(A,B)) == NULL) return FALSE

/** 
 * Binary read function with byte swaping (assume file is BIG ENDIAN)
 * 
 * @param fp [in] file pointer
 * @param buf [out] read data
 * @param unitbyte [in] size of a unit in bytes
 * @param unitnum [in] number of unit to be read
 */
static boolean
rdnfunc(FILE *fp, void *buf, size_t unitbyte, int unitnum)
{
  size_t tmp;

  if (unitnum == 0) return TRUE;

  if (gzfile) {
    tmp = myfread(buf, unitbyte, unitnum, fp);
  } else {
    tmp = fread(buf, unitbyte, unitnum, fp);
  }
  if (tmp < (size_t)unitnum) {
    jlog("Error: read_binhmm: failed to read %d bytes\n", unitbyte * unitnum);
    return FALSE;
  }
#ifndef WORDS_BIGENDIAN
  if (unitbyte != 1) {
    swap_bytes(buf, unitbyte, unitnum);
  }
#endif
  return TRUE;
}

static char buf[MAXLINELEN];	///< Local work are for text handling
static char nostr = '\0';
/** 
 * Read a string till NULL.
 * 
 * @param fp [in] file pointer
 * @param hmm [out] pointer to %HMM definition data to store the values.
 * 
 * @return pointer to a newly allocated buffer that contains the read string.
 */
static char *
rdn_strfunc(FILE *fp, HTK_HMM_INFO *hmm)
{
  int c;
  int len;
  char *p;

  len = 0;
  while ((c = gzfile ? myfgetc(fp) : fgetc(fp)) != -1) {
    if (len >= MAXLINELEN) {
      jlog("Error: read_binhmm: string len exceeded %d bytes\n", MAXLINELEN);
      jlog("Error: read_binhmm: please check the value of MAXLINELEN\n");
      return NULL;
    }
    buf[len++] = c;
    if (c == '\0') break;
  }
  if (len == 0) return NULL;
  if (len == 1) {
    p = &nostr;
  } else {
    p = (char *)mybmalloc2(len, &(hmm->mroot));
    strcpy(p, buf);
  }
  return(p);
}


static char *binhmm_header = BINHMM_HEADER; ///< Header string
static char *binhmm_header_v2 = BINHMM_HEADER_V2; ///< Header string for V2

/** 
 * Read acoustic analysis configration parameters from header of binary HMM.
 * 
 * @param fp [in] file pointer
 * @param para [out] acoustic analysis configration parameters
 */
static boolean
rd_para(FILE *fp, Value *para)
{
  short version;
  float dummy;

  /* read version */
  rdn(fp, &version, sizeof(short), 1);

  if (version > VALUE_VERSION) {
    jlog("Error: read_binhmm: unknown embedded parameter format version: %d\n", version);
    return FALSE;
  }
  jlog("Stat: rd_para: found embedded acoutic parameter (ver.%d)\n", version);

  /* read parameters */
  rdn(fp, &(para->smp_period), sizeof(long), 1);      
  rdn(fp, &(para->smp_freq), sizeof(long), 1);	
  rdn(fp, &(para->framesize), sizeof(int), 1);        
  rdn(fp, &(para->frameshift), sizeof(int), 1);       
  rdn(fp, &(para->preEmph), sizeof(float), 1);        
  rdn(fp, &(para->lifter), sizeof(int), 1);           
  rdn(fp, &(para->fbank_num), sizeof(int), 1);        
  rdn(fp, &(para->delWin), sizeof(int), 1);           
  rdn(fp, &(para->accWin), sizeof(int), 1);           
  rdn(fp, &(para->silFloor), sizeof(float), 1);       
  rdn(fp, &(para->escale), sizeof(float), 1);         
  rdn(fp, &(para->hipass), sizeof(int), 1);		
  rdn(fp, &(para->lopass), sizeof(int), 1);		
  rdn(fp, &(para->enormal), sizeof(int), 1);          
  rdn(fp, &(para->raw_e), sizeof(int), 1);            
  if (version == 1) {
    /* version 1 has ss related parameters, but version 2 and later not */
    /* skip ss related parameters (ss_alpha and ss_floor) */
    rdn(fp, &dummy, sizeof(float), 1);
    rdn(fp, &dummy, sizeof(float), 1);
  }
  rdn(fp, &(para->zmeanframe), sizeof(int), 1);	
  if (version >= 3) {
    rdn(fp, &(para->usepower), sizeof(int), 1);
  }

  return(TRUE);
}

/** 
 * Read header string of binary HMM file.
 * 
 * @param fp [in] file pointer
 * @param hmm [out] pointer to %HMM definition data to store the values.
 * @param para [out] store embedded acoustic parameters if any (V2)
 * @param mpdf_macro_ret [out] will be set to TRUE if the file contains mixture pdf macro defined by "~p"
 * 
 * @return TRUE if a correct header was read, FALSE if header string does not
 * match the current version.
 */
static boolean
rd_header(FILE *fp, HTK_HMM_INFO *hmm, Value *para, boolean *mpdf_macro_ret)
{
  char *p, *q;
  boolean emp, inv;
  
  rdn_str(fp, hmm, p);
  if (strmatch(p, binhmm_header)) {
    /* version 1 */
    hmm->variance_inversed = FALSE;
  } else if (strmatch(p, binhmm_header_v2)) {
    /* version 2 */
    emp = inv = FALSE;
    rdn_str(fp, hmm, q);
    if (*q != '\0') {
      while(*q == '_') {
	q++;
	switch (*q) {
	case BINHMM_HEADER_V2_EMBEDPARA:
	  /* read in embedded acoutic condition parameters */
	  emp = TRUE;
	  jlog("Stat: binhmm-header: analysis parameter embedded\n");
	  break;
	case BINHMM_HEADER_V2_VARINV:
	  inv = TRUE;
	  jlog("Stat: binhmm-header: variance inversed\n");
	  break;
	case BINHMM_HEADER_V2_MPDFMACRO:
	  *mpdf_macro_ret = TRUE;
	  jlog("Stat: binhmm-header: mixture PDF macro used\n");
	  break;
	default:
	  jlog("Error: unknown format qualifier in header: \"%c\"\n", *q);
	  return FALSE;
	}
	q++;
      }
    }
    if (emp) {
      para->loaded = 1;
      if (rd_para(fp, para) == FALSE) {
	jlog("Error: read_binhmm: failed to read embeded parameter\n");
	return FALSE;
      }
      jlog("Stat: read_binhmm: has acoutic analysis configurations in its header\n");
    }
    if (inv) {
      hmm->variance_inversed = TRUE;
      jlog("Stat: read_binhmm: has inversed variances\n");
    } else {
      hmm->variance_inversed = FALSE;
    }
  } else {
    /* failed to read header */
    return FALSE;
  }
  return TRUE;
}



/** 
 * Read %HMM option specifications.
 * 
 * @param fp [in] file pointer
 * @param opt [out] pointer to the %HMM option structure to hold the read
 * values.
 */
static boolean
rd_opt(FILE *fp, HTK_HMM_Options *opt)
{
  rdn(fp, &(opt->stream_info.num), sizeof(short), 1);
  rdn(fp, opt->stream_info.vsize, sizeof(short), MAXSTREAMNUM);
  rdn(fp, &(opt->vec_size), sizeof(short), 1);
  rdn(fp, &(opt->cov_type), sizeof(short), 1);
  rdn(fp, &(opt->dur_type), sizeof(short), 1);
  rdn(fp, &(opt->param_type), sizeof(short), 1);

  return(TRUE);
}

/** 
 * Read %HMM type of mixture tying.
 * 
 * @param fp [in] file pointer
 * @param hmm [out] pointer to %HMM definition data to store the values.
 */
static boolean
rd_type(FILE *fp, HTK_HMM_INFO *hmm)
{
  rdn(fp, &(hmm->is_tied_mixture), sizeof(boolean), 1);
  rdn(fp, &(hmm->maxmixturenum), sizeof(int), 1);
  return TRUE;
}


/* read transition data */
static HTK_HMM_Trans **tr_index; ///< Map transition matrix id to its pointer
static unsigned int tr_num;	///< Length of above

/** 
 * @brief  Read a sequence of transition matrix data for @a tr_num.
 *
 * The transition matrixes are stored into @a hmm, and their pointers
 * are also stored in @a tr_index for later data mapping operation
 * from upper structure (state etc.).
 * 
 * @param fp [in] file pointer
 * @param hmm [out] %HMM definition structure to hold the read transitions.
 */
static boolean
rd_trans(FILE *fp, HTK_HMM_INFO *hmm)
{
  HTK_HMM_Trans *t;
  unsigned int idx;
  int i;
  PROB *atmp;
  char *p;

  rdn(fp, &tr_num, sizeof(unsigned int), 1);
  tr_index = (HTK_HMM_Trans **)mymalloc(sizeof(HTK_HMM_Trans *) * tr_num);

  hmm->trstart = NULL;
  hmm->tr_root = NULL;
  for (idx = 0; idx < tr_num; idx++) {
    t = (HTK_HMM_Trans *)mybmalloc2(sizeof(HTK_HMM_Trans), &(hmm->mroot));
    rdn_str(fp, hmm, p);
    t->name = (*p == '\0') ? NULL : p;
    rdn(fp, &(t->statenum), sizeof(short), 1);
    t->a = (PROB **)mybmalloc2(sizeof(PROB *) * t->statenum, &(hmm->mroot));
    atmp = (PROB *)mybmalloc2(sizeof(PROB) * t->statenum * t->statenum, &(hmm->mroot));
    for (i=0;i<t->statenum;i++) {
      t->a[i] = &(atmp[i*t->statenum]);
      rdn(fp, t->a[i], sizeof(PROB), t->statenum);
    }
    trans_add(hmm, t);
    tr_index[idx] = t;
  }

#ifdef DMES
  jlog("Stat: read_binhmm: %d transition maxtix read\n", tr_num);
#endif
  return TRUE;
}


static HTK_HMM_Var **vr_index;	///< Map variance id to its pointer
static unsigned int vr_num;	///< Length of above

/** 
 * @brief  Read a sequence of variance vector for @a vr_num.
 *
 * The variance vectors are stored into @a hmm, and their pointers
 * are also stored in @a vr_index for later data mapping operation
 * from upper structure (density etc.).
 * 
 * @param fp [in] file pointer
 * @param hmm [out] %HMM definition structure to hold the read variance.
 */
static boolean
rd_var(FILE *fp, HTK_HMM_INFO *hmm)
{
  HTK_HMM_Var *v;
  unsigned int idx;
  char *p;

  rdn(fp, &vr_num, sizeof(unsigned int), 1);
  vr_index = (HTK_HMM_Var **)mymalloc(sizeof(HTK_HMM_Var *) * vr_num);
  
  hmm->vrstart = NULL;
  hmm->vr_root = NULL;
  for (idx = 0; idx < vr_num; idx++) {
    v = (HTK_HMM_Var *)mybmalloc2(sizeof(HTK_HMM_Var), &(hmm->mroot));
    rdn_str(fp, hmm, p);
    v->name = (*p == '\0') ? NULL : p;
    rdn(fp, &(v->len), sizeof(short), 1);
    v->vec = (VECT *)mybmalloc2(sizeof(VECT) * v->len, &(hmm->mroot));
    rdn(fp, v->vec, sizeof(VECT), v->len);
    vr_index[idx] = v;
    var_add(hmm, v);
  }
#ifdef DMES
  jlog("Stat: read_binhmm: %d variance read\n", vr_num);
#endif
  return TRUE;
}


/* read density data */
static HTK_HMM_Dens **dens_index; ///< Map density id to its pointer
static unsigned int dens_num;	///< Length of above

/** 
 * @brief  Read a sequence of mixture densities for @a dens_num.
 *
 * The mixture densities are stored into @a hmm, and their references
 * to lower structure (variance etc.) are recovered from the id-to-pointer
 * index.  Their pointers are also stored in @a dens_index for
 * later data mapping operation from upper structure (state etc.).
 * 
 * @param fp [in] file pointer
 * @param hmm [out] %HMM definition structure to hold the read densities.
 */
static boolean
rd_dens(FILE *fp, HTK_HMM_INFO *hmm)
{
  HTK_HMM_Dens *d;
  unsigned int idx;
  unsigned int vid;
  char *p;

  rdn(fp, &dens_num, sizeof(unsigned int), 1);
  hmm->totalmixnum = dens_num;
  dens_index = (HTK_HMM_Dens **)mymalloc(sizeof(HTK_HMM_Dens *) * dens_num);

  hmm->dnstart = NULL;
  hmm->dn_root = NULL;
  for (idx = 0; idx < dens_num; idx++) {
    d = (HTK_HMM_Dens *)mybmalloc2(sizeof(HTK_HMM_Dens), &(hmm->mroot));
    rdn_str(fp, hmm, p);
    d->name = (*p == '\0') ? NULL : p;
    rdn(fp, &(d->meanlen), sizeof(short), 1);
    d->mean = (VECT *)mybmalloc2(sizeof(VECT) * d->meanlen, &(hmm->mroot));
    rdn(fp, d->mean, sizeof(VECT), d->meanlen);
    rdn(fp, &vid, sizeof(unsigned int), 1);
    d->var = vr_index[vid];
    rdn(fp, &(d->gconst), sizeof(LOGPROB), 1);
    dens_index[idx] = d;
    dens_add(hmm, d);
  }
#ifdef DMES
  jlog("Stat: read_binhmm: %d gaussian densities read\n", dens_num);
#endif
  return TRUE;
}


/* read stream weight data */
static HTK_HMM_StreamWeight **streamweight_index; ///< Map stream weights id to its pointer
static unsigned int streamweight_num;	///< Length of above

/** 
 * @brief  Read a sequence of stream weights for @a streamweight_num.
 *
 * The stream weights are stored into @a hmm, and their references
 * to lower structure (variance etc.) are recovered from the id-to-pointer
 * index.  Their pointers are also stored in @a dens_index for
 * later data mapping operation from upper structure (state etc.).
 * 
 * @param fp [in] file pointer
 * @param hmm [out] %HMM definition structure to hold the read stream weights.
 */
static boolean
rd_streamweight(FILE *fp, HTK_HMM_INFO *hmm)
{
  HTK_HMM_StreamWeight *sw;
  unsigned int idx;
  char *p;