g++ [flags ...] file ... -l /isip/tools/lib/$ISIP_BINARY/lib_search.a

#include <HierarchicalSearch.h>

HierarchicalSearch();
HierarchicalSearch(const HierarchicalSearch& copy_search);
boolean setNumLevels(long num_levels);
SearchLevel& getSearchLevel(long level)
boolean decode(FrontEnd& fe, long num_frames = DEF_NUM_FRAMES);
boolean getHypotheses(String& output_hyp, float& total_score,
                      long& num_frames, long level = 0,
                      boolean detailed = false,
                      boolean cumulative = false);

// configure FrontEnd
//
FrontEnd fe;

Sof front_end_sof;
front_end_sof.open(L"front_end.sof");
fe.read(front_end_sof, front_end_sof.first(fe.name()));
front_end_sof.close();

// configure HierarchicalSearch
//
HierarchicalSearch search_engine;
long num_levels = 3;
search_engine.setNumLevels(num_levels);

// read models
//
Sof model_sof;
model_sof.open(L"models.sof");

for (long curr_level = 0; curr_level < num_levels; curr_level++) {
  SearchLevel& level = search_engine.getSearchLevel(curr_level);
  level.load(model_sof);
}
model_sof.close();

// process the sentence
//
Filename speech_file(L"speech_file.raw");
fe.open(speech_file);
search_engine.decode(fe);
fe.close();

// get the best hypothesis
//
String best_hypothesis;
float score = 0;
long num_frames = 0;
search_engine.getHypotheses(best_hypothesis, score, num_frames);

• The search space is hierarchical in that each level has nodes and each node is either:
  • composed of a sub-graph + probability evaluation
  • composed of probability evaluation.
• At each level you will be able to do the following:
  • beam prune
  • instance prune
  • generate context-dependent scoring models on the fly including cross-symbol models
  • specify whether search information is to be maintained - otherwise we can prune away redundant information. For instance, once a word is completed, it is no longer necessary to maintain the sub-word units that correspond to that word.
  • turn on/off sub-level evaluation
• Each graph position will be capable of the following:
  • generate a probability of being in that graph position given the input stream. For acoustic models, this will involve direct evaluation of the data stream. For levels like language models, this will not involve manipulation of the data stream at all. Any level can access the data stream as it will be passed from the highest level all the way down to the lowest via pointer reference.
  • store the overlapping paths in the sub-graph as a "lexical" tree.
  • use N-symbol models related to the possible entries on that level
• We will revolve around the concept of a "trace" where the trace moves through the graph and keeps track of the paths/scores through the search space

• none.

• none.