m_info.c

julius version 4.12.about sound recognition.

    jlog("\n\n");

    if (r->wchmm != NULL) {
      print_wchmm_info(r->wchmm); jlog("\n");
    }
    if (r->lmtype == LM_PROB) {
      jlog(" Inter-word N-gram cache: \n");
      {
	int num, len;
#ifdef UNIGRAM_FACTORING
	len = r->wchmm->isolatenum;
	jlog("\troot node to be cached = %d / %d (isolated only)\n",
	     len, r->wchmm->startnum);
#else
	len = r->wchmm->startnum;
	jlog("\troot node to be cached = %d (all)\n", len);
#endif
#ifdef HASH_CACHE_IW
	num = (r->config->pass1.iw_cache_rate * r->lm->ngram->max_word_num) / 100;
	jlog("\tword ends to be cached = %d / %d\n", num, r->lm->ngram->max_word_num);
#else
	num = r->lm->ngram->max_word_num;
	jlog("\tword ends to be cached = %d (all)\n", num);
#endif
	jlog("\t  max. allocation size = %dMB\n", num * len / 1000 * sizeof(LOGPROB) / 1000);
      }
    }

    if (r->lmtype == LM_PROB) {
      jlog("\t(-lmp)  pass1 LM weight = %2.1f  ins. penalty = %+2.1f\n", r->config->lmp.lm_weight, r->config->lmp.lm_penalty);
      jlog("\t(-lmp2) pass2 LM weight = %2.1f  ins. penalty = %+2.1f\n", r->config->lmp.lm_weight2, r->config->lmp.lm_penalty2);
      jlog("\t(-transp)trans. penalty = %+2.1f per word\n", r->config->lmp.lm_penalty_trans);
    } else if (r->lmtype == LM_DFA && r->lmvar == LM_DFA_GRAMMAR) {
      jlog("\t(-penalty1) IW penalty1 = %+2.1f\n", r->config->lmp.penalty1);
      jlog("\t(-penalty2) IW penalty2 = %+2.1f\n", r->config->lmp.penalty2);
    }


#ifdef CONFIDENCE_MEASURE
#ifdef CM_MULTIPLE_ALPHA
    jlog("\t(-cmalpha)CM alpha coef = from %f to %f by step of %f (%d outputs)\n", r->config->annotate.cm_alpha_bgn, r->config->annotate.cm_alpha_end, r->config->annotate.cm_alpha_step, r->config->annotate.cm_alpha_num);
#else
    jlog("\t(-cmalpha)CM alpha coef = %f\n", r->config->annotate.cm_alpha);
#endif
#ifdef CM_SEARCH_LIMIT
    jlog("\t(-cmthres) CM cut thres = %f for hypo generation\n", r->config->annotate.cm_cut_thres);
#endif
#ifdef CM_SEARCH_LIMIT_POP
    jlog("\t(-cmthres2)CM cut thres = %f for popped hypo\n", r->config->annotate.cm_cut_thres_pop);
#endif
#endif /* CONFIDENCE_MEASURE */
    jlog("\n");

    if (r->am->hmminfo->multipath) {
      if (r->lm->config->enable_iwsp) {
	jlog("\t inter-word short pause = on (append \"%s\" for each word tail)\n", r->am->hmminfo->sp->name);
	jlog("\t  sp transition penalty = %+2.1f\n", r->am->config->iwsp_penalty);
      }
    }

    if (r->lmvar == LM_DFA_WORD) {
#ifdef DETERMINE
      jlog("    early word determination:  (-wed)\n");
      jlog("\tscore threshold    = %f\n", r->config->pass1.determine_score_thres);
      jlog("\tframe dur. thres   = %d\n", r->config->pass1.determine_duration_thres);
#endif
    }

    jlog(" Search parameters: \n");
    jlog("\t    multi-path handling = ");
    if (r->am->hmminfo->multipath) {
      jlog("yes, multi-path mode enabled\n");
    } else {
      jlog("no\n");
    }
    jlog("\t(-b) trellis beam width = %d", r->trellis_beam_width);
    if (r->config->pass1.specified_trellis_beam_width == -1) {
      jlog(" (-1 or not specified - guessed)\n");
    } else if (r->config->pass1.specified_trellis_beam_width == 0) {
      jlog(" (0 - full)\n");
    } else {
      jlog("\n");
    }
    jlog("\t(-n)search candidate num= %d\n", r->config->pass2.nbest);
    jlog("\t(-s)  search stack size = %d\n", r->config->pass2.stack_size);
    jlog("\t(-m)    search overflow = after %d hypothesis poped\n", r->config->pass2.hypo_overflow);
    jlog("\t        2nd pass method = ");
    if (r->config->graph.enabled) {
#ifdef GRAPHOUT_DYNAMIC
#ifdef GRAPHOUT_SEARCH
      jlog("searching graph, generating dynamic graph\n");
#else
      jlog("searching sentence, generating dynamic graph\n");
#endif /* GRAPHOUT_SEARCH */
#else  /* ~GRAPHOUT_DYNAMIC */
      jlog("searching sentence, generating static graph from N-best\n");
#endif
    } else {
      jlog("searching sentence, generating N-best\n");
    }
    if (r->config->pass2.enveloped_bestfirst_width >= 0) {
      jlog("\t(-b2)  pass2 beam width = %d\n", r->config->pass2.enveloped_bestfirst_width);
    }
    jlog("\t(-lookuprange)lookup range= %d  (tm-%d <= t <tm+%d)\n",r->config->pass2.lookup_range,r->config->pass2.lookup_range,r->config->pass2.lookup_range);
#ifdef SCAN_BEAM
    jlog("\t(-sb)2nd scan beamthres = %.1f (in logscore)\n", r->config->pass2.scan_beam_thres);
#endif
    jlog("\t(-n)        search till = %d candidates found\n", r->config->pass2.nbest);
    jlog("\t(-output)    and output = %d candidates out of above\n", r->config->output.output_hypo_maxnum);

    if (r->ccd_flag) {
      jlog("\t IWCD handling:\n");
#ifdef PASS1_IWCD
      jlog("\t   1st pass: approximation ");
      switch(r->am->hmminfo->cdset_method) {
      case IWCD_AVG:
	jlog("(use average prob. of same LC)\n");
	break;
      case IWCD_MAX:
	jlog("(use max. prob. of same LC)\n");
	break;
      case IWCD_NBEST:
	jlog("(use %d-best of same LC)\n", r->am->hmminfo->cdmax_num);
	break;
      }
#else
      jlog("\t   1st pass: ignored\n");
#endif
#ifdef PASS2_STRICT_IWCD
      jlog("\t   2nd pass: strict (apply when expanding hypo. )\n");
#else
      jlog("\t   2nd pass: loose (apply when hypo. is popped and scanned)\n");
#endif
    }
    if (r->lmtype == LM_PROB) {
      jlog("\t factoring score: ");
#ifdef UNIGRAM_FACTORING
      jlog("1-gram prob. (statically assigned beforehand)\n");
#else
      jlog("2-gram prob. (dynamically computed while search)\n");
#endif
    }

    if (r->config->annotate.align_result_word_flag) {
      jlog("\t output word alignments\n");
    }
    if (r->config->annotate.align_result_phoneme_flag) {
      jlog("\t output phoneme alignments\n");
    }
    if (r->config->annotate.align_result_state_flag) {
      jlog("\t output state alignments\n");
    }
    if (r->lmtype == LM_DFA && r->lmvar == LM_DFA_GRAMMAR) {
      if (r->config->pass2.looktrellis_flag) {
	jlog("\t only words in backtrellis will be expanded in 2nd pass\n");
      } else {
	jlog("\t all possible words will be expanded in 2nd pass\n");
      }
    }
    if (r->wchmm != NULL) {
      if (r->wchmm->category_tree) {
	if (r->config->pass1.old_tree_function_flag) {
	  jlog("\t build_wchmm() used\n");
	} else {
	  jlog("\t build_wchmm2() used\n");
	}
#ifdef PASS1_IWCD
#ifdef USE_OLD_IWCD
	jlog("\t full lcdset used\n");
#else
	jlog("\t lcdset limited by word-pair constraint\n");
#endif
#endif /* PASS1_IWCD */
      }
    }
    if (r->config->output.progout_flag) {
      jlog("\tprogressive output on 1st pass\n");
    }
    if (r->config->compute_only_1pass) {
      jlog("\tCompute only 1-pass\n");
    }
    
    if (r->config->graph.enabled) {
      jlog("\n");
      jlog("Graph-based output with graph-oriented search:\n");
      jlog("\t(-lattice)      word lattice = %s\n", r->config->graph.lattice ? "yes" : "no");
      jlog("\t(-confnet) confusion network = %s\n", r->config->graph.confnet ? "yes" : "no");
      if (r->config->graph.lattice == TRUE) {
	jlog("\t(-graphrange)         margin = %d frames", r->config->graph.graph_merge_neighbor_range);
	if (r->config->graph.graph_merge_neighbor_range < 0) {
	  jlog(" (all post-marging disabled)\n");
	} else if (r->config->graph.graph_merge_neighbor_range == 0) {
	  jlog(" (merge same word with the same boundary)\n");
	} else {
	  jlog(" (merge same words around this margin)\n");
	}
      }
#ifdef GRAPHOUT_DEPTHCUT
      jlog("\t(-graphcut)cutoff depth      = ");
      if (r->config->graph.graphout_cut_depth < 0) {
	jlog("disabled (-1)\n");
      } else {
	jlog("%d words\n",r->config->graph.graphout_cut_depth);
      }
#endif
#ifdef GRAPHOUT_LIMIT_BOUNDARY_LOOP
      jlog("\t(-graphboundloop)loopmax     = %d for boundary adjustment\n",r->config->graph.graphout_limit_boundary_loop_num);
#endif
#ifdef GRAPHOUT_SEARCH_DELAY_TERMINATION
      jlog("\tInhibit graph search termination before 1st sentence found = ");
      if (r->config->graph.graphout_search_delay) {
	jlog("enabled\n");
      } else {
	jlog("disabled\n");
      }
#endif

    }
    
    if (r->config->successive.enabled) {
      jlog("\tshort pause segmentation = on\n");
      jlog("\t      sp duration length = %d frames\n", r->config->successive.sp_frame_duration);
#ifdef SPSEGMENT_NAIST
      jlog("      backstep margin on trigger = %d frames\n", r->config->successive.sp_margin);
      jlog("\t        delay on trigger = %d frames\n", r->config->successive.sp_delay);
#endif
      if (r->config->successive.pausemodelname) {
	jlog("\t   pause models for seg. = %s\n",  r->config->successive.pausemodelname);
      }
    } else {
      jlog("\tshort pause segmentation = off\n");
    }
    if (r->config->output.progout_flag) {
      jlog("\t        progout interval = %d msec\n", r->config->output.progout_interval);
    }
    jlog("\tfall back on search fail = ");
    if (r->config->sw.fallback_pass1_flag) {
      jlog("on, adopt 1st pass result as final\n");
    } else {
      jlog("off, returns search failure\n");
    }

    jlog("\n");
  }

  jlog("------------------------------------------------------------\n");
  jlog("Decoding algorithm:\n\n");
  jlog("\t1st pass input processing = ");
  if (jconf->decodeopt.force_realtime_flag) jlog("(forced) ");
  if (jconf->decodeopt.realtime_flag) {
    jlog("real time, on-the-fly\n");
  } else {
    jlog("buffered, batch\n");
  }
  jlog("\t1st pass method = ");
#ifdef WPAIR
# ifdef WPAIR_KEEP_NLIMIT
  jlog("word-pair approx., keeping only N tokens ");
# else
  jlog("word-pair approx. ");
# endif
#else
  jlog("1-best approx. ");
#endif
#ifdef WORD_GRAPH
  jlog("generating word_graph\n");
#else
  jlog("generating indexed trellis\n");
#endif
#ifdef CONFIDENCE_MEASURE
  jlog("\toutput word confidence measure ");
#ifdef CM_NBEST
  jlog("based on N-best candidates\n");
#endif
#ifdef CM_SEARCH
  jlog("based on search-time scores\n");
#endif
#endif /* CONFIDENCE_MEASURE */
  
  jlog("\n");

  jlog("------------------------------------------------------------\n");
  jlog("FrontEnd:\n\n");

  jlog(" Input stream:\n");
  jlog("\t             input type = ");
  switch(jconf->input.type) {
  case INPUT_WAVEFORM:
    jlog("waveform\n");
    break;
  case INPUT_VECTOR:
    jlog("feature vector sequence\n");
    break;
  }
  jlog("\t           input source = ");
  if (jconf->input.plugin_source != -1) {
    jlog("plugin\n");
  } else if (jconf->input.speech_input == SP_RAWFILE) {
    jlog("waveform file\n");
    jlog("\t          input filelist = ");
    if (jconf->input.inputlist_filename == NULL) {
      jlog("(none, get file name from stdin)\n");
    } else {
      jlog("%s\n", jconf->input.inputlist_filename);
    }
  } else if (jconf->input.speech_input == SP_MFCFILE) {
    jlog("feature vector file (HTK format)\n");
    jlog("\t                filelist = ");
    if (jconf->input.inputlist_filename == NULL) {
      jlog("(none, get file name from stdin)\n");
    } else {
      jlog("%s\n", jconf->input.inputlist_filename);
    }
  } else if (jconf->input.speech_input == SP_STDIN) {
    jlog("standard input\n");
  } else if (jconf->input.speech_input == SP_ADINNET) {
    jlog("adinnet client\n");
#ifdef USE_NETAUDIO
  } else if (jconf->input.speech_input == SP_NETAUDIO) {
    char *p;
    jlog("NetAudio server on ");
    if (jconf->input.netaudio_devname != NULL) {
      jlog("%s\n", jconf->input.netaudio_devname);
    } else if ((p = getenv("AUDIO_DEVICE")) != NULL) {
      jlog("%s\n", p);
    } else {
      jlog("local port\n");
    }
#endif
  } else if (jconf->input.speech_input == SP_MIC) {
    jlog("microphone\n");
    jlog("\t    device API          = ");
    switch(jconf->input.device) {
    case SP_INPUT_DEFAULT: jlog("default\n"); break;
    case SP_INPUT_ALSA: jlog("alsa\n"); break;
    case SP_INPUT_OSS: jlog("oss\n"); break;
    case SP_INPUT_ESD: jlog("esd\n"); break;
    }
  }
  if (jconf->input.type == INPUT_WAVEFORM) {
    if (jconf->input.speech_input == SP_RAWFILE || jconf->input.speech_input == SP_STDIN || jconf->input.speech_input == SP_ADINNET) {
      if (jconf->input.use_ds48to16) {
	jlog("\t          sampling freq. = assume 48000Hz, then down to %dHz\n", jconf->input.sfreq);
      } else {
	jlog("\t          sampling freq. = %d Hz required\n", jconf->input.sfreq);
      }
    } else {
      if (jconf->input.use_ds48to16) {
	jlog("\t          sampling freq. = 48000Hz, then down to %d Hz\n", jconf->input.sfreq);
      } else {
 	jlog("\t          sampling freq. = %d Hz\n", jconf->input.sfreq);
      }
    }
  }
  if (jconf->input.type == INPUT_WAVEFORM) {
    jlog("\t         threaded A/D-in = ");
#ifdef HAVE_PTHREAD
    if (recog->adin->enable_thread) {
      jlog("supported, on\n");
    } else {
      jlog("supported, off\n");
    }
#else
    jlog("not supported (live input may be dropped)\n");
#endif
  }
  if (jconf->preprocess.strip_zero_sample) {
    jlog("\t   zero frames stripping = on\n");
  } else {
    jlog("\t   zero frames stripping = off\n");
  }
  if (jconf->input.type == INPUT_WAVEFORM) {
    if (recog->adin->adin_cut_on) {
      jlog("\t         silence cutting = on\n");
      jlog("\t             level thres = %d / 32767\n", jconf->detect.level_thres);
      jlog("\t         zerocross thres = %d / sec.\n", jconf->detect.zero_cross_num);
      jlog("\t             head margin = %d msec.\n", jconf->detect.head_margin_msec);
      jlog("\t             tail margin = %d msec.\n", jconf->detect.tail_margin_msec);
    } else {
      jlog("\t         silence cutting = off\n");
    }

    if (jconf->preprocess.use_zmean) {
      jlog("\t    long-term DC removal = on");
      if (jconf->input.speech_input == SP_RAWFILE) {
	jlog(" (will compute for each file)\n");
      } else {
	jlog(" (will compute from first %.1f sec)\n", (float)ZMEANSAMPLES / (float)jconf->input.sfreq);
      }
    } else {
      jlog("\t    long-term DC removal = off\n");
    }
  }
  jlog("\t      reject short input = ");
  if (jconf->reject.rejectshortlen > 0) {
    jlog("< %d msec\n", jconf->reject.rejectshortlen);
  } else {
    jlog("off\n");
  }
#ifdef POWER_REJECT
  jlog("\t   power rejection thres = %f", jconf->reject.powerthres);
#endif

  jlog("\n");

  jlog("----------------------- System Information end -----------------------\n");

#ifdef USE_MIC
  if (jconf->decodeopt.realtime_flag) {
    boolean flag;
    flag = FALSE;
    for(mfcc=recog->mfcclist; mfcc; mfcc=mfcc->next) {
      if (mfcc->para->cmn && mfcc->cmn.loaded) {
	flag = TRUE;
	break;
      }
    }
    if (flag) {
      jlog("\n");
      jlog("initial CMN parameter loaded from file\nfor");
      for(mfcc=recog->mfcclist; mfcc; mfcc=mfcc->next) {
	if (mfcc->para->cmn && mfcc->cmn.loaded) {
	  jlog(" MFCC%02d", mfcc->id);
	}
      }
      jlog("\n");
    }
    flag = FALSE;
    for(mfcc=recog->mfcclist; mfcc; mfcc=mfcc->next) {
      if (mfcc->para->cmn && !mfcc->cmn.loaded) {
	flag = TRUE;
	break;
      }
    }
    if (flag) {
      jlog("\n");
      jlog("\t*************************************************************\n");
      jlog("\t* NOTICE: The first input may not be recognized, since      *\n");
      jlog("\t*         no initial CMN parameter is available on startup. *\n");
      jlog("\t* for");
      for(mfcc=recog->mfcclist; mfcc; mfcc=mfcc->next) {
	if (mfcc->para->cmn && !mfcc->cmn.loaded) {
	  jlog(" MFCC%02d", mfcc->id);
	}
      }
      jlog("*\n");
      jlog("\t*************************************************************\n");
    }
    flag = FALSE;
    for(mfcc=recog->mfcclist; mfcc; mfcc=mfcc->next) {
      if (mfcc->para->energy && mfcc->para->enormal) {
	flag = TRUE;
	break;
      }
    }
    if (flag) {
      jlog("\t*************************************************************\n");
      jlog("\t* NOTICE: Energy normalization is activated on live input:  *\n");
      jlog("\t*         maximum energy of LAST INPUT will be used for it. *\n");
      jlog("\t*         So, the first input will not be recognized.       *\n");
      jlog("\t* for");
      for(mfcc=recog->mfcclist; mfcc; mfcc=mfcc->next) {
	if (mfcc->para->energy && mfcc->para->enormal) {
	  jlog(" MFCC%02d", mfcc->id);
	}
      }
      jlog("*\n");
      jlog("\t*************************************************************\n");
    }
  }
#endif
}

/* end of file */