dec_decode_3.cc

这是处理语音信号的程序

// file: dec_decode_3.cc
//

// system include files
//
#include <unistd.h>

// isip include files
//
#include "decoder.h"
#include "decoder_constants.h"
#include <time.h>

// method: decode_cc
//
// arguments:
//  FILE* fin: (input) feature input file
//  FILE* fout: (input) decoder output file
//  FILE* flat: (input) decoder output lattice file
//  FILE* fnbest: (input) decoder output nbest file
//
// return: logical_1 indicating status
//
// this method decodes an utterance specified by the input file and
// writes the output to the output file, and the generated lattice ---
// if any --- to the lattice output file
//
logical_1 Decoder::decode_cc(FILE* fin_a, FILE* fout_a, FILE* flat_a,
			     FILE* nbest_a) {

  // time counting variables
  //
  float_8 begin, end;

  // start the counter
  //
  decode_time_d = 0;
  begin = (float_8) clock();
  
  // reset the decoder to prepare for decoding this utterance
  //
  reset_cc();

  // print the initial list of word-hypotheses to file
  //
  if (demo_mode_d == ISIP_TRUE) {
    write_output_cc(fout_a);
  }
  num_steps_d = (int_4)0;

  // flag indicating the start of utterrance
  //
  logical_1 start = ISIP_TRUE;

  // read HTK header
  //
  if (input_feature_format_d == DEC_HTK_FORMAT) {
    read_htk_header_cc(fin_a);
  }
  
  // loop over all frames of input feature data
  //
  while (read_features_cc(fin_a, start) == ISIP_TRUE) {
        
    // increment the current time frame
    //
    frame_d++;

    // reset counts
    //
    for (int_4 i = 0; i < num_levels_d; i++) {
      num_traces_gen_d[i] = (int_4)0;
      num_traces_del_d[i] = (int_4)0;
    }    
      
    // if this is the demo mode
    //
    if (demo_mode_d == ISIP_TRUE) {
      
      // read number of steps to advance frames
      //
      read_key_cc();
      
      // if this is the abort number, quit loop and go to find
      // sentence end hyps
      //
      if (num_steps_d < (int_4)0) {
	break;
      }
    }
    
    // loop over all active word and phone-level traces in the
    // current time frame and grow the next phone
    //
    enter_models_cc();

    // maximum active model instance pruning on all instances of
    // models propagated from previous frame
    //
    prune_inst_cc();
    
    // loop over all active state-level traces
    // evaluate the state and update the score for each trace
    // then find all possible transitions and advance traces
    //
    project_states_cc();

    // find the beam pruning score for state-level pruning
    //
    set_beam_cc(DEC_STATE_LEVEL);
    
    // prune away any traces that do not satisfy the new state beam
    //
    beam_prune_cc();
    
    // find the phone level beam pruning threshold
    //
    set_beam_cc(DEC_MODEL_LEVEL);

    // create word-level traces for the word-end phone-level traces
    // make sure that the inactive traces are no longer kept
    // 
    exit_word_cc();

    // find the word level beam pruning threshold
    //
    set_beam_cc(DEC_WORD_LEVEL);

    // sort the word score list to get the cutoff score
    //
    word_cutoff_cc();
    
    // store counts
    //
    for (int_4 i = 0; i < num_levels_d; i++) {
      total_gen_d[i] += num_traces_gen_d[i];
      total_del_d[i] += num_traces_del_d[i];
      num_traces_total_d[i] += (num_traces_gen_d[i] - num_traces_del_d[i]);
    }

    // output the frame status
    //
    if (frame_d % DEC_FRAMES_OUTNUM == (int_4)0) {

      // get the end counter after decoding 100 frames
      // 'else' counts on the time when wrapping around
      //
      end = (float_8) clock();
      if (begin <= end) {
	decode_time_d += (double) (end - begin);
      }
      else {
	decode_time_d += (2147483647 + end) + (2147483647 - begin);
      }
     
      fprintf(stdout, "Processed %ld frames: decoding time = %10.1f seconds, %7.2f xRT\n", frame_d, decode_time_d/1000000, decode_time_d/(frame_d*10000));
      fflush(stdout);

      // start the counter
      //
      begin = (float_8) clock();
    }

    // dump the frame statistics to output
    //
    if ((demo_mode_d == ISIP_TRUE) && (num_steps_d == (int_4)0)) {
      
      // the word-level traces are used to determine the best word
      // hypotheses at the end of this frame
      //
      find_nbest_cc();
      
      // print the n-best list of word-hypotheses to file
      //
      write_output_cc(fout_a);

    } // end if demo mode
  } // end while loop over all frames
    
  // get the end counter after decoding all the frames
  // 'else' counts on the time when wrapping around
  //
  end = (float_8) clock();
  if (begin <= end) {
    decode_time_d += (double) (end - begin);
  }
  else {
    decode_time_d += (2147483647 + end) + (2147483647 - begin);
  }
  
  fprintf(stdout, "Processed %ld frames: decoding time = %10.1f seconds, %7.2f xRT\n", frame_d, decode_time_d/1000000, decode_time_d/(frame_d*10000));
  fflush(stdout);
  
  // start the counter
  //
  begin = (float_8) clock();
  
  // we have now finished the state sequences so we need to backtrace
  // i.e. find an emitting word level trace with the highest score and
  // follow backpointers to the start trace
  //
  backtrace_cc();
    
  // if no sentence end is found print the best word end anyway
  //
  if (num_hyps_d == 0) {
    
    // print error message and find the best word end
    //
    fprintf(stdout, "No sentence-end hypotheses were found.\n");
    find_nbest_cc();
  }

  // print the n-best list of sentence-hypotheses to file
  //
  write_output_cc(fout_a);
  
  // if this is lattice generation mode build a lattice from this,
  // only if there are valid paths
  //
  if ((num_hyps_d > 0) &&
      ((function_mode_d == DEC_LATTICE_GENERATE_FUNCTION) ||
       (function_mode_d == DEC_LATTICE_LATTICE_FUNCTION))) {
    build_lattice_cc(flat_a);
  }

  // print the n-best list of sentence-hypotheses to file
  //  
  else if ((num_nbest_d > 1) && (nbest_a != NULL)) {
    Lattice* lat = NULL;
    build_lattice_cc(lat);
    lattice_to_nbest_cc(lat, nbest_a, nbest_max_paths_d, nbest_beam_d);
    delete lat;
  }
  
  // clean up memory used for this utterance
  //
  clean_cc();

  // get the end counter after decoding all the frames
  // 'else' counts on the time when wrapping around
  //
  end = (float_8) clock();
  if (begin <= end) {
    decode_time_d += (float_8) (end - begin);
  }
  else {
    decode_time_d += (2147483647 + end) + (2147483647 - begin);
  }
  
  fprintf(stdout, "Decoded the utterance: %ld frames, %.1f seconds, %.2f xRT\n", frame_d, decode_time_d/1000000, decode_time_d/(frame_d*10000));
  fflush(stdout);

  // update the overall counter
  //
  decode_time_d /= 1000000;
  
  // line break
  //
  fprintf(stdout, "\n");

  // exit gracefully
  //
  return ISIP_TRUE;
}