1 /*---------------------------------------------------------------------------* 2 * srec_initialize.c * 3 * * 4 * Copyright 2007, 2008 Nuance Communciations, Inc. * 5 * * 6 * Licensed under the Apache License, Version 2.0 (the 'License'); * 7 * you may not use this file except in compliance with the License. * 8 * * 9 * You may obtain a copy of the License at * 10 * http://www.apache.org/licenses/LICENSE-2.0 * 11 * * 12 * Unless required by applicable law or agreed to in writing, software * 13 * distributed under the License is distributed on an 'AS IS' BASIS, * 14 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * 15 * See the License for the specific language governing permissions and * 16 * limitations under the License. * 17 * * 18 *---------------------------------------------------------------------------*/ 19 20 #ifndef _RTT 21 #include "pstdio.h" 22 #endif 23 #include <stdlib.h> 24 #include <string.h> 25 #include <math.h> 26 #include "passert.h" 27 28 #include "portable.h" 29 30 #include "hmm_desc.h" 31 #include "utteranc.h" 32 #include "hmmlib.h" 33 34 #include "srec_sizes.h" 35 #include "srec.h" 36 #include "word_lattice.h" 37 #include "swimodel.h" 38 39 #include "c42mul.h" 40 41 /*this file contains code which handles the initialization of the srec data structures*/ 42 43 /*allocates an srec - 44 45 input args come from config and are: 46 47 int viterbi_prune_thresh; score-based pruning threshold - only keep paths within this delta of best cost 48 49 int max_hmm_tokens; controls the maximum number of HMM's alive in any frame. If number 50 exceeded, pruning gets tightened. So, this threshold can be used 51 to tradeoff accuracy for computation an memory 52 int max_fsmnode_tokens; controls the maximum number of FSMs alive in any frame. If number, 53 exceeded, pruning gets tightened. So, this threshold can be used 54 to tradeoff accuracy for computation an memory 55 int max_word_tokens; controls the maximum number of word tokens kept in the word lattice. 56 if number exceeded, the word lattice is pruned more tightly (less word 57 ends per frame 58 59 int max_altword_tokens; controls the maximum number of alternative paths to propagate for proper nbest 60 61 int num_wordends_per_frame; controls the size of the word lattice - the number of word ends to 62 keep at each time frame 63 int max_fsm_nodes; allocation size of a few arrays in the search - needs to be big enough 64 to handle any grammar that the search needs to run. Initialization fails 65 if num exceeded 66 int max_fsm_arcs; allocation size of a few arrays in the search - needs to be big enough 67 to handle any grammar that the search needs to run. Initialization fails 68 if num exceeded 69 70 */ 71 72 static void allocate_recognition1(srec *rec, 73 int viterbi_prune_thresh, /*score-based pruning threshold - only keep paths within this delta of best cost*/ 74 int max_hmm_tokens, 75 int max_fsmnode_tokens, 76 int max_word_tokens, 77 int max_altword_tokens, 78 int num_wordends_per_frame, 79 int max_frames, 80 int max_model_states) 81 { 82 #ifdef SREC_ENGINE_VERBOSE_LOGGING 83 PLogMessage("allocating recognition arrays2 prune %d max_hmm_tokens %d max_fsmnode_tokens %d max_word_tokens %d max_altword_tokens %d max_wordends_per_frame %d\n", 84 viterbi_prune_thresh, 85 max_hmm_tokens, 86 max_fsmnode_tokens, 87 max_word_tokens, 88 max_altword_tokens, 89 num_wordends_per_frame); 90 #endif 91 rec->current_model_scores = (costdata*) CALLOC_CLR(max_model_states, sizeof(costdata), "search.srec.current_model_scores"); /*FIX - either get NUM_MODELS from acoustic models, or check this someplace to make sure we have enough room*/ 92 rec->num_model_slots_allocated = (modelID)max_model_states; 93 94 rec->fsmarc_token_array_size = (stokenID)max_hmm_tokens; 95 96 rec->fsmarc_token_array = (fsmarc_token*) CALLOC_CLR(rec->fsmarc_token_array_size , sizeof(fsmarc_token), "search.srec.fsmarc_token_array"); 97 rec->max_new_states = (stokenID)max_hmm_tokens; 98 99 rec->word_token_array = (word_token*) CALLOC_CLR(max_word_tokens, sizeof(word_token), "search.srec.word_token_array"); 100 rec->word_token_array_size = (wtokenID)max_word_tokens; 101 /* todo: change this to a bit array later */ 102 rec->word_token_array_flags = (asr_int16_t*) CALLOC_CLR(max_word_tokens, sizeof(asr_int16_t), "search.srec.word_token_array_flags"); 103 104 rec->fsmnode_token_array = (fsmnode_token*) CALLOC_CLR(max_fsmnode_tokens, sizeof(fsmnode_token), "search.srec.fsmnode_token_array"); 105 rec->fsmnode_token_array_size = (ftokenID)max_fsmnode_tokens; 106 107 rec->altword_token_array = (altword_token*) CALLOC_CLR(max_altword_tokens, sizeof(altword_token), "search.srec.altword_token_array"); 108 rec->altword_token_array_size = (wtokenID)max_altword_tokens; 109 110 rec->prune_delta = (costdata)viterbi_prune_thresh; 111 112 rec->max_frames = (frameID)max_frames; 113 rec->best_model_cost_for_frame = (costdata*)CALLOC_CLR(max_frames, sizeof(costdata), "search.srec.best_model_cost_for_frame"); 114 rec->word_lattice = allocate_word_lattice((frameID)max_frames); 115 116 rec->word_priority_q = allocate_priority_q(num_wordends_per_frame); 117 rec->best_fsmarc_token = MAXstokenID; 118 119 #define ASTAR_NBEST_LEN 10 120 rec->astar_stack = astar_stack_make(rec, ASTAR_NBEST_LEN); 121 rec->context = NULL; 122 } 123 124 static int check_parameter_range(int parval, int parmin, int parmax, const char* parname) 125 { 126 if (parval > parmax) 127 { 128 log_report("Error: %s value %d is out-of-range [%d,%d]\n", parname, 129 parval, parmin, parmax); 130 return 1; 131 } 132 else 133 { 134 return 0; 135 } 136 } 137 138 int allocate_recognition(multi_srec *rec, 139 int viterbi_prune_thresh, /*score-based pruning threshold - only keep paths within this delta of best cost*/ 140 int max_hmm_tokens, 141 int max_fsmnode_tokens, 142 int max_word_tokens, 143 int max_altword_tokens, 144 int num_wordends_per_frame, 145 int max_fsm_nodes, 146 int max_fsm_arcs, 147 int max_frames, 148 int max_model_states, 149 int max_searches) 150 { 151 int i; 152 153 if (check_parameter_range(max_fsm_nodes, 1, MAXnodeID, "max_fsm_nodes")) 154 return 1; 155 if (check_parameter_range(max_fsm_arcs, 1, MAXarcID, "max_fsm_arcs")) 156 return 1; 157 if (check_parameter_range(max_frames, 1, MAXframeID, "max_frames")) 158 return 1; 159 if (check_parameter_range(max_model_states, 1, MAXmodelID, "max_model_states")) 160 return 1; 161 if (check_parameter_range(max_hmm_tokens, 1, MAXstokenID, "max_hmm_tokens")) 162 return 1; 163 if (check_parameter_range(max_fsmnode_tokens, 1, MAXftokenID, "max_fsmnode_tokens")) 164 return 1; 165 if (check_parameter_range(viterbi_prune_thresh, 1, MAXcostdata, "viterbi_prune_thresh")) 166 return 1; 167 if (check_parameter_range(max_altword_tokens, 0, MAXftokenID, "max_altword_tokens")) 168 return 1; 169 if (check_parameter_range(max_searches, 1, 2, "max_searches")) 170 return 1; 171 172 rec->rec = (srec*)CALLOC_CLR(max_searches, sizeof(srec), "search.srec.base"); 173 rec->num_allocated_recs = max_searches; 174 rec->num_swimodels = 0; 175 176 /* best_token_for_arc and best_token_for_node are shared across 177 multiple searches */ 178 rec->best_token_for_arc = (stokenID*)CALLOC_CLR(max_fsm_arcs, sizeof(stokenID), "search.srec.best_token_for_arc"); 179 rec->max_fsm_arcs = (arcID)max_fsm_arcs; 180 181 rec->best_token_for_node = (ftokenID*)CALLOC_CLR(max_fsm_nodes, sizeof(ftokenID), "search.srec.best_token_for_node"); 182 rec->max_fsm_nodes = (nodeID)max_fsm_nodes; 183 184 /* cost offsets and accumulated cost offsets are pooled for all 185 different searches, this saves memory and enables each search 186 to know it's total scores */ 187 rec->cost_offset_for_frame = (costdata*)CALLOC_CLR(max_frames, sizeof(costdata), "search.srec.current_best_costs"); 188 rec->accumulated_cost_offset = (bigcostdata*)CALLOC_CLR(max_frames, sizeof(bigcostdata), "search.srec.accumulated_cost_offset"); 189 rec->max_frames = (frameID)max_frames; 190 for (i = 0; i < max_frames; i++) 191 rec->accumulated_cost_offset[i] = 0; 192 193 /* now copy the shared data down to individual recogs */ 194 for (i = 0; i < rec->num_allocated_recs; i++) 195 { 196 allocate_recognition1(&rec->rec[i], viterbi_prune_thresh, max_hmm_tokens, max_fsmnode_tokens, max_word_tokens, max_altword_tokens, num_wordends_per_frame, max_frames, max_model_states); 197 rec->rec[i].best_token_for_node = rec->best_token_for_node; 198 rec->rec[i].max_fsm_nodes = rec->max_fsm_nodes; 199 rec->rec[i].best_token_for_arc = rec->best_token_for_arc; 200 rec->rec[i].max_fsm_arcs = rec->max_fsm_arcs; 201 rec->rec[i].max_frames = rec->max_frames; 202 rec->rec[i].cost_offset_for_frame = rec->cost_offset_for_frame; 203 rec->rec[i].accumulated_cost_offset = rec->accumulated_cost_offset; 204 rec->rec[i].id = (asr_int16_t)i; 205 } 206 rec->eos_status = VALID_SPEECH_NOT_YET_DETECTED; 207 return 0; 208 } 209 210 211 static void free_recognition1(srec *rec) 212 { 213 FREE(rec->current_model_scores); 214 FREE(rec->fsmarc_token_array); 215 FREE(rec->word_token_array); 216 FREE(rec->word_token_array_flags); 217 FREE(rec->fsmnode_token_array); 218 FREE(rec->altword_token_array); 219 FREE(rec->best_model_cost_for_frame); 220 destroy_word_lattice(rec->word_lattice); 221 free_priority_q(rec->word_priority_q); 222 astar_stack_destroy(rec); 223 } 224 225 void free_recognition(multi_srec *rec) 226 { 227 int i; 228 for (i = 0; i < rec->num_allocated_recs; i++) 229 free_recognition1(&rec->rec[i]); 230 FREE(rec->accumulated_cost_offset); 231 FREE(rec->cost_offset_for_frame); 232 FREE(rec->best_token_for_node); 233 FREE(rec->best_token_for_arc); 234 FREE(rec->rec); 235 } 236 237
