[#1354] Merge r75 changes from cm_mods branch to trunk to implement a penalty for repeated non-silence HMM loop transitions.

git-svn-id: https://gforge.ti.com/svn/tiesr/trunk@76 469fd053-429d-4955-9e4c-93c0d7a8ff27

2 files changed, 115 insertions, 61 deletions

diff --git a/TIesrEngine/src/gmhmm.h b/TIesrEngine/src/gmhmm.h
index d88909f..45b76e0 100644
--- a/TIesrEngine/src/gmhmm.h
+++ b/TIesrEngine/src/gmhmm.h
@@ -78,7 +78,8 @@ Now in tiesrcommonmacros.h
 
 #define PRUNE_COEFF -3  /* default pruning */
 #define WIP 0   /*  default word insertion penalty */
-
+#define TRAN_DUR_BIAS  -100
+//#define TRAN_DUR_BIAS 0
 #define NSYM 2 /* hash code nbr */
 
 //#define   MAX_STATE_CELL     20000    /* these 4 const define search space */
@@ -106,7 +107,7 @@ Now in tiesrcommonmacros.h
 
 /* From search.cpp */
 #define UNDEF_CELL  0xffff
-#define UNDEFINEDTIME (-1)
+#define UNDEFINEDTIME 0xfff
 
 
 /* ---------------------------------------------------------------------------
@@ -249,6 +250,10 @@ typedef enum CompensationType {
       (time_idx) ? ((base) + (pos)) : \
       ((base) + max_state_cell - (pos) - n_state)
 
+#define   STATE_FRM( frm )  ( frm & 0xfff)
+#define   STATE_DUR( frm )  ( (frm >> 12) & 0xf )
+#define   STATE_CAT( dur, frm )    ( ((dur & 0xf) << 12 ) | (frm & 0xfff) )
+
 /* was: ((base) + MAX_STATE_CELL - 1 - (pos) - n_state) */
 
 #define   SYM_CELL(base,pos,time_idx,max_sym_cell2) \
diff --git a/TIesrEngine/src/search.cpp b/TIesrEngine/src/search.cpp
index e23b93b..e3d3c0a 100755
--- a/TIesrEngine/src/search.cpp
+++ b/TIesrEngine/src/search.cpp
@@ -374,7 +374,7 @@ static void coloring_beam(unsigned short cur_time_index, gmhmm_type *gv)
   unsigned short q, b_index, cont, c_bm_cnt = 0; 
   short depth; 
 
-  if (cur_time_index == USHRT_MAX) return ; /* no beam cell has this time (no sym was collected at this time) */
+  if (cur_time_index == UNDEFINEDTIME) return ; /* no beam cell has this time (no sym was collected at this time) */
   for (q =  gv->time_base[cur_time_index].beam_idx; NOT_END_IX_CELL(q,gv->max_beam_cell); q = beam_base[q].next) { 
     bm = beam_base + q;
     depth=0;
@@ -574,7 +574,7 @@ grow_beam_cell(unsigned short sym, StateCell *p_cell, short first_time, unsigned
   /* previous frame sym, starting frame, last frame*/
   
   gv->beam_base[idx].idx_flag = (p_cell->sym == USHRT_MAX)?  gv->max_beam_cell  //zlook
-          : find_beam_index(p_cell->sym, p_cell->frm, gv, last_frm);  //previous stats' symble and starting frame.
+          : find_beam_index(p_cell->sym, STATE_FRM(p_cell->frm), gv, last_frm);  //previous stats' symble and starting frame.
 
 
 #ifdef _MONOPHONE_NET
@@ -587,7 +587,7 @@ grow_beam_cell(unsigned short sym, StateCell *p_cell, short first_time, unsigned
         
     gv->beam_count=0;
     gv->next_time = next_time_cell(gv);
-    if (gv->next_time == USHRT_MAX) 
+    if (gv->next_time == USHRT_MAX)
       return eTIesrEngineTimeCellOverflow;
     gv->time_base[gv->next_time].frm = frm_cnt;
     gv->time_base[gv->next_time].beam_idx = idx; /* this was changed from address to index */
@@ -724,13 +724,14 @@ static void init_search_last_syms( gmhmm_type * gv)
    ---------------------------------------------------------------------- */
 
 static void
-within_model_path(HmmType *hmm,  StateCell *p_prev,  StateCell *p_cur, gmhmm_type *gv, short prune)
+within_model_path(unsigned short hmmcode, HmmType *hmm,  StateCell *p_prev,  StateCell *p_cur, gmhmm_type *gv, short prune)
 {
   long       cur_prob;
   short      *trnprob;
   short      i, j;
   StateCell  *state_cell, *pprev;
   short      n_state, n1;
+  short      tprob, dur;
 
   n_state = NBR_STATES(hmm,gv->base_tran);
   trnprob = gv->base_tran + hmm->tran + n_state;
@@ -740,18 +741,49 @@ within_model_path(HmmType *hmm,  StateCell *p_prev,  StateCell *p_cur, gmhmm_typ
 
     if ( pprev->scr > prune ) {  /*if a possible start -- within model prune*/
       
-      for (j = 0; j < n1; j++) {   /*for all target states within model*/
-
-        if ( trnprob[j] > BAD_SCR ) {          /* there exist i to j jump. note transprob is updated above with i */
+      /*for all target states within model*/
+      for (j = 0; j < n1; j++) {   
+
+       /* Add duration penalty to transition probability 
+          if not silence HMM and is a loop transition */
+       if(  hmmcode != 0 && i == j )
+       {
+          dur = STATE_DUR(pprev->frm);
+          tprob = LONG2SHORT( (long)trnprob[j] + (long)dur*TRAN_DUR_BIAS );
+       }
+       else
+       {
+          tprob = trnprob[j];
+       }
+         
+        /* there exist i to j jump. note transprob is updated above with i */
+        if ( tprob > BAD_SCR ) {          
 
           state_cell = p_cur + j;
-          cur_prob = (long) pprev->scr + (long) trnprob[j];
+          cur_prob = (long) pprev->scr + (long) tprob;
           
-          if ( cur_prob > (long) state_cell->scr ) {  //viterbi on transition prob.
+          /* Viterbi decoding done here on transition probability */
+          if ( cur_prob > (long) state_cell->scr ) {  
             state_cell->scr = (short) cur_prob; 
-            state_cell->sym = pprev->sym;   /* propogate prev sym & frm */
-            state_cell->frm = pprev->frm;  /*starting time, B*/
+            
+            /* propogate previous symbol */
+            state_cell->sym = pprev->sym; 
+            
+            /* Propagate previous symbol frame, and state duration */
+            if( hmmcode != 0  &&  i == j )
+            {
+               /* loop transition in non-silence HMM */
+               dur++;
+               dur = MIN( dur, 15 );
+               state_cell->frm = STATE_CAT( dur, pprev->frm );
+            }
+            else
+            {
+               /* Set last frame count only, which clears duration count */
+               state_cell->frm = STATE_FRM(pprev->frm);  
+            }
           }
+          
         }
       }
     }
@@ -842,14 +874,18 @@ cross_model_path(short exit_scr, unsigned short v_sym, unsigned short t_sym,
                           start_scr >= (long) start_cell->scr, gv); 
         }
 #endif
-        if ( start_scr >= (long) start_cell->scr ) { 
-          /* Viterbi, for the winner,
-             remember the frame index and the symbol where we are from */
-          start_cell->scr = (short) start_scr;
-          start_cell->frm = starttime; //frm_cnt - 1;  // this is true for HMM backtrace. not word based backtrace.
-          start_cell->sym = v_sym;  // word sym
-
+        
+        /* Viterbi decoding is done here in transitioning from the exit state
+         of the prior symbol's HMM into a state of a next grammar symbol HMM.
+         Store the frame index and the symbol that started evaluation of this
+         next HMM. Depending on the type of search (WORD or HMM) the symbol
+         that started evaluation may be the prior symbol or prior
+         word-ending symbol. */
+        if ( start_scr >= (long) start_cell->scr ) {
 
+          start_cell->scr = (short) start_scr;
+          start_cell->frm = STATE_FRM(starttime); 
+          start_cell->sym = v_sym;  
         }
 #ifdef USE_NBEST
         if (gv->bUseNBest) nb_uniq_nbest(gv->ptr_nbest_tmp->pNBestCells, uBeamIdx, start_cell, gv); 
@@ -945,7 +981,7 @@ path_propagation(short cur_idx, short prev_idx, unsigned short frm_cnt, gmhmm_ty
        }
 
        /*label all the within model path, if score is OK, prune is OK, */
-       within_model_path(hmm, p_prev, p_cur, gv, prune); /*hmm, last state cell, new allocated state cell, ...*/
+       within_model_path(hmmcode, hmm, p_prev, p_cur, gv, prune); /*hmm, last state cell, new allocated state cell, ...*/
         
        exit_prob = p_prev[ n_state - 1 ].scr;
 
@@ -973,7 +1009,7 @@ path_propagation(short cur_idx, short prev_idx, unsigned short frm_cnt, gmhmm_ty
 
          }
          else     // p_prev->sym will be the word info of last sym
-         { starttime = p_prev[ n_state - 1 ].frm;
+         { starttime = STATE_FRM(p_prev[ n_state - 1 ].frm);
            status = cross_model_path(exit_prob + gv->tranwgt, p_prev[n_state -1].sym, t_sym, hmm_dlt, sym_dlt, cur_idx, frm_cnt,gv, sym_prev->depth,starttime);
          
           if ( status != eTIesrEngineSuccess )  
@@ -1035,15 +1071,16 @@ static short
 
   long           scr;
   StateCell      *max_exit_state = NULL, *state_cell, *start_cell;
-  unsigned short last_SYM = USHRT_MAX, last_FRM = USHRT_MAX;
+  unsigned short last_SYM = USHRT_MAX, last_FRM = UNDEFINEDTIME;
   unsigned short bestdepth = 0, n_actsym=0,sumdepth=0, maxdepth=0;
+  unsigned short frm;
   
 #ifdef SHOW_DIAGONISIS
   if (sttnode!=3) sttnode=0;
 #endif
 
 #ifdef USE_CONFIDENCE
-  short *p_scr_anti, *pScrAnti = cm_antiscr((ConfType*)gv->pConf, frm_cnt); 
+  short *p_scr_anti, *pScrAnti = cm_antiscr((ConfType*)gv->pConf, frm_cnt);
   short     /* for cache */
     cm_prev_scr = LZERO; 
   if (((ConfType*)gv->pConf)->bNeedUpdate == FALSE) 
@@ -1131,19 +1168,19 @@ static short
 
 #ifdef USE_CONFIDENCE
           if ((gs > cm_prev_scr) && (pScrAnti != NULL)){
-            p_scr_anti =  pScrAnti; 
+            p_scr_anti =  pScrAnti;
 
             if (p_scr_anti){
-              *p_scr_anti = gs; 
+              *p_scr_anti = gs;
             }
-            
-            cm_prev_scr = gs; 
-          }
+
+            cm_prev_scr = gs;
+            }
 #endif
 
-         scr = (long) state_cell->scr + (long) gs; //gauss_obs_score_f(sObs?sObs:mfcc_feature, pdf_idx, gv); 
+         scr = (long) state_cell->scr + (long) gs; //gauss_obs_score_f(sObs?sObs:mfcc_feature, pdf_idx, gv);
+
 
-           
         //if (sym_cell->depth *1.5  - gv->frm_cnt>0) scr-= 500;  // soft prune
         //if ( sym_cell->depth > (gv->bestdepth +9) || sym_cell->depth < (gv->bestdepth -10)  ) scr-=500;
           scr -= (long) gv->best_prev_scr;
@@ -1154,14 +1191,15 @@ static short
             /* coloring pending states (including exit states):
              * mark active beam cells for active pending states
              */
-            if (state_cell->frm != USHRT_MAX) 
-              if ((state_cell->frm != last_FRM) || (state_cell->sym != last_SYM)) {  //if it is from a new search node (frame/sym pair)
+            frm = STATE_FRM( state_cell->frm );
+            if ( frm != UNDEFINEDTIME ) 
+              if ( (frm != last_FRM) || (state_cell->sym != last_SYM)) {  //if it is from a new search node (frame/sym pair)
                 /*pass best frame/sym to this state so far*/
                 if (gv->need_compact==TRUE ) //|| gv->word_backtrace==FALSE) //if (gv->evalstat.beam.cur_pos > 1000)   //if (gv->frm_cnt==14)
                   {   
-                        mark_cells_state(state_cell->frm, state_cell->sym, gv, last_frm);   
+                        mark_cells_state( frm, state_cell->sym, gv, last_frm);   
                   }   
-                last_FRM = state_cell->frm; last_SYM = state_cell->sym; 
+                last_FRM = frm; last_SYM = state_cell->sym; 
               }
           }
 
@@ -1175,11 +1213,13 @@ static short
 #ifdef MODEL_LEVEL_PRUNE
           mlp_update(gv, hmmcode , scr);
 #endif
-        
-          if ( *trnprob > BAD_SCR ) {  /* exit? */
+          /* Check if there is a transition to the exit state. */
+          if ( *trnprob > BAD_SCR ) {  
 
             scr = (long) state_cell->scr  +  (long) *trnprob;
-          
+
+            /* Viterbi decode here for the highest likelihood state transition
+             to the exit state.*/
             if ( scr > (long) max_exit_scr) { 
               max_exit_scr = (short) scr;
               max_exit_state = state_cell;
@@ -1192,11 +1232,14 @@ static short
       else { /* exit state */
         sumdepth+=sym_cell->depth;  n_actsym++;      
         maxdepth=sym_cell->depth> maxdepth?sym_cell->depth: maxdepth;
-            
+
+        /* If a valid transition to the exit state exists, update the exit
+         state cell with the information corresponding to the highest
+         likelihood exit transition. */
         if ( max_exit_scr > BAD_SCR ) {
           state_cell->scr = max_exit_scr;
           state_cell->sym = max_exit_state->sym;
-          state_cell->frm = max_exit_state->frm;
+          state_cell->frm = STATE_FRM(max_exit_state->frm);
         
           /*if end of sym, put it in the beam cell list*/
           // word end label, grow beam cell
@@ -1257,7 +1300,7 @@ init_search(short mfcc_feature[], short cur_idx, gmhmm_type *gv, unsigned short
   short          status, n_state, pdf_idx, *startprob, *trnprob, first_time_call = TRUE;
   EvalIdx  *eval_idx = gv->eval+cur_idx;
 
-  
+
 #ifdef SHOW_DIAGONISIS
   sttnode=0;
   global_count=0;
@@ -1309,10 +1352,13 @@ init_search(short mfcc_feature[], short cur_idx, gmhmm_type *gv, unsigned short
         return eTIesrEngineEvalCellOverflow;
       }
 
-      for (j = 0; j < n_state - 1; j++) {          /* for all emit states . */
-
-        if ( startprob[ j ] > BAD_SCR ) {          /* locate inititial stateS */
+      /* For all emitting states */
+      for (j = 0; j < n_state - 1; j++) {          
+        
+        /* Check if state has initial entry probability */
+        if ( startprob[ j ] > BAD_SCR ) {          
 
+          /* Determine initial state likelihood score */
           pdf_idx = hmm->pdf_idx[ j ];
 
         /*  if (gv->obs_scr[ pdf_idx ] == 0 )  // a cache in computing. 
@@ -1329,15 +1375,16 @@ init_search(short mfcc_feature[], short cur_idx, gmhmm_type *gv, unsigned short
           else state_cell[ j ].scr = (short) scr;
 
           if ( scr > (long) gv->best_cur_scr ) 
-            {
+          {
             gv->best_cur_scr = (short) scr;  /*first frame only*/
-            
-            }
+
+            }
 
 #ifdef MODEL_LEVEL_PRUNE
           mlp_update(gv, hmmcode , scr);
 #endif
-
+          /* If a transition from state to exit state exists, determine its
+           likelihood.*/
           if ( trnprob[ j * n_state + exit_state ] > BAD_SCR ) {
             scr = (long) state_cell[ j ].scr + 
               (long) trnprob[ j * n_state + exit_state ];
@@ -1346,16 +1393,19 @@ init_search(short mfcc_feature[], short cur_idx, gmhmm_type *gv, unsigned short
               max_exit_scr = (short) scr;   /*track the best score to exit the sym*/
           }
         }
-        state_cell[ j ].frm = USHRT_MAX;     /*the 4 lines below say : there is no prior sym and frame*/
+        
+        /* Set prior symbol and frame for the state to invalid values indicating
+         beginning of search. */
+        state_cell[ j ].frm = UNDEFINEDTIME;     
         state_cell[ j ].sym = USHRT_MAX;
       } /* end of all emiting states */
 
-      state_cell[ exit_state ].frm = USHRT_MAX; 
+      /* Set prior symbol and frame of exit state to invalid values */
+      state_cell[ exit_state ].frm = UNDEFINEDTIME; 
       state_cell[ exit_state ].sym = USHRT_MAX;
 
-      /* HMM (sym) completed, create a backtrace cell */
-
-      if ( max_exit_scr > BAD_SCR ) {   /* if can be end of symbol, but sure not for first frame. so why here ? */
+       /* If valid likelihood of exit, update exit state likelihood score. */
+      if ( max_exit_scr > BAD_SCR ) {   
         state_cell[ exit_state ].scr = max_exit_scr;
         
         
@@ -1363,19 +1413,18 @@ init_search(short mfcc_feature[], short cur_idx, gmhmm_type *gv, unsigned short
         //status = grow_beam_cell(VTL_SYM(t_symcode, sym_dlt), state_cell + exit_state, first_time_call, 0, 
         //                      gv, last_frm);
         gv->ending_sym_count ++;
-        if (psym->n_next & 0x8000 || gv->word_backtrace!=WORDBT)  //never should happen in word bt
-          {
-                   status = grow_beam_cell(VTL_SYM(t_symcode, sym_dlt), state_cell + exit_state, first_time_call, 0, 
+        
+        /* Add the symbol/prior info of the exit state to the beam cell linked
+         list for this frame if HMM is word ending or if HMM backtrace requested */
+        if (psym->n_next & 0x8000 || gv->word_backtrace!=WORDBT) 
+        {
+            status = grow_beam_cell(VTL_SYM(t_symcode, sym_dlt), state_cell + exit_state, first_time_call, 0, 
                                 gv, last_frm);
                    first_time_call = FALSE;   //first time to call beam cell stuff -- for this frame!
              gv->beam_count++;
              if ( status != eTIesrEngineSuccess ) 
                return status;     
-          }
-        
-        
-        
-        
+        }
       }
     }
   }