jconf.h

来自「julius version 4.12.about sound recognit」· C头文件 代码 · 共 1,202 行 · 第 1/2 页

/**
 * @file   jconf.h
 * 
 * <JA>
 * @brief  Jconf 菇陇挛の年盗
 *
 * </JA>
 * 
 * <EN>
 * @brief  Jconf structure
 *
 * </EN>
 * 
 * <pre>
 * JCONF
 *   +- JCONF_AM[] (linked list)
 *   +- JCONF_LM[] (linked list)
 *   +- JCONF_SEARCH[] (linked list) -> each has pointer to *JCONF_AM, *JCONF_LM
 *   +- JCONF_AM for GMM
 *   +- (engine configurations)
 * </pre>
 * 
 * @author Akinobu Lee
 * @date   Fri Feb 16 13:42:28 2007
 *
 * $Revision: 1.8 $
 * 
 */
/*
 * Copyright (c) 1991-2007 Kawahara Lab., Kyoto University
 * Copyright (c) 2000-2005 Shikano Lab., Nara Institute of Science and Technology
 * Copyright (c) 2005-2007 Julius project team, Nagoya Institute of Technology
 * All rights reserved
 */

/*
*/

#ifndef __J_JCONF_H__
#define __J_JCONF_H__

#include <sent/stddefs.h>
#include <sent/hmm.h>
#include <sent/vocabulary.h>
#include <sent/ngram2.h>
#include <sent/dfa.h>
#include <sent/mfcc.h>
#include <julius/wchmm.h>
#include <julius/search.h>
#include <julius/callback.h>

/**
 * Maximum length of module name string
 * 
 */
#define JCONF_MODULENAME_MAXLEN 64

/**
 * Default module name if not specified (in case of old jconf)
 * 
 */
#define JCONF_MODULENAME_DEFAULT "_default"

/**
 * Configuration parameters (AM)
 * 
 */
typedef struct __jconf_am__ {

  /**
   * Unique ID
   * 
   */
  short id;

  /**
   * Unique name
   * 
   */
  char name[JCONF_MODULENAME_MAXLEN];

  /**
   * HMM definition file (-h)
   */
  char *hmmfilename;
  /**
   * HMMList file to map logical (tri)phones to physical models (-hlist)
   */
  char *mapfilename;
  /**
   * Gaussian pruning method (-gprune)
   * Default: use value from compile-time engine configuration default.
   */
  int gprune_method;
  /**
   * Number of Gaussian to compute per mixture on Gaussian pruning (-tmix)
     */
  int mixnum_thres;   
  /**
   * Logical HMM name of short pause model (-spmodel)
   * Default: "sp"
   */
  char *spmodel_name;
  /**
   * GMS: HMM definition file for GMS (-gshmm)
   */
  char *hmm_gs_filename;
  /**
   * GMS: number of mixture PDF to select (-gsnum)
   */
  int gs_statenum;    

  /**
   * Calculation method for outprob score of a lcdset on cross-word
   * triphone (-iwcd1) 
   */
  short iwcdmethod;
  
  /**
   * N-best states to be calculated on IWCD_NBEST (-iwcd1 best N)
   */
  short iwcdmaxn;

  /**
   * Transition penalty of inter-word short pause (-iwsppenalty)
   * for multi-path mode
   */
  LOGPROB iwsp_penalty; 

  /**
   * force multipath mode
   * 
   */
  boolean force_multipath;

  /**
   * Acoustic Analysis Conditions.  Parameter setting priority is:
   * user-specified > specified HTK Config > model-embedded > Julius default.
   * 
   */
  struct {
    /**
     * All MFCC computation parameters, actually used for recognition.
     */
    Value para;         
    /**
     * default parameters of Julius
     */
    Value para_default;
    /**
     * parameters from binhmm header
     */
    Value para_hmm;             
    /**
     * parameters from HTK Config (-htkconf)
     */
    Value para_htk;     
    /**
     * CMN: load initial cepstral mean from file at startup (-cmnload)
     */
    char *cmnload_filename;
    /**
     * CMN: update cepstral mean while recognition
     * (-cmnnoupdate to unset)
     */
    boolean cmn_update;
    /**
     * CMN: save cepstral mean to file at end of every recognition (-cmnsave)
     */
    char *cmnsave_filename;     
    /**
     * CMN: MAP weight for initial cepstral mean on (-cmnmapweight)
     */
    float cmn_map_weight;

  } analysis;


  /**
   * Frontend processing parameter for this AM
   * 
   */
  struct {
    /**
     * Alpha coefficient for spectral subtraction
     * 
     */
    float ss_alpha;

    /**
     * Flooring coefficient for spectral subtraction
     * 
     */
    float ss_floor;

    /**
     * SS: compute noise spectrum from head silence on file input (-sscalc)
     */
    boolean sscalc;

    /**
     * With "-sscalc", specify noise length at input head in msec (-sscalclen)
     */
    int sscalc_len;

    /**
     * Load noise spectrum data from file (-ssload), that was made by "mkss".
     */
    char *ssload_filename;
  } frontend;

  /**
   * Plugin source ID when using plugin (gprune_method is GPRUNE_SEL_USER)
   */
  int gprune_plugin_source;

  /* pointer to next instance */
  struct __jconf_am__ *next;

} JCONF_AM;

/**
 * Language models (N-gram / DFA), dictionary, and related parameters.
 * 
 */
typedef struct __jconf_lm__ {

  /**
   * Unique ID
   * 
   */
  short id;

  /**
   * Unique name
   * 
   */
  char name[JCONF_MODULENAME_MAXLEN];

  /**
   * Language model type: one of LM_UNDEF, LM_NGRAM, LM_DFA
   * 
   */
  int lmtype;

  /**
   * Variation type of language model: one of LM_NGRAM, LM_DFA_GRAMMAR,
   * LM_DFA_WORD
   * 
   */
  int lmvar;

  /**
   * Word dictionary file (-v)
   */
  char *dictfilename;
  
  /**
   * Silence word to be placed at beginning of speech (-silhead) for N-gram
   */
  char *head_silname;
  /**
   * Silence word to be placed at end of search (-siltail) for N-gram
   */
  char *tail_silname;
  
  /**
   * Skip error words in dictionary and continue (-forcedict)
   */
  boolean forcedict_flag;
  
  /**
   * N-gram in binary format (-d)
   */
  char *ngram_filename;
  /**
   * LR 2-gram in ARPA format (-nlr)
   */
  char *ngram_filename_lr_arpa;
  /**
   * RL 3-gram in ARPA format (-nrl)
   */
  char *ngram_filename_rl_arpa;
  
  /**
   * DFA grammar file (-dfa, for single use)
   */
  char *dfa_filename;
  
  /**
   * List of grammars to be read at startup (-gram) (-gramlist)
   */
  GRAMLIST *gramlist_root;
  
  /**
   * List of word lists to be read at startup (-w) (-wlist)
   */
  GRAMLIST *wordlist_root;
  
  /**
   * Enable inter-word short pause handling on multi-path version (-iwsp)
   * for multi-path mode
   */
  boolean enable_iwsp; 
  
  /**
   * Enable automatic addition of "short pause word" to the dictionary
   * (-iwspword) for N-gram
   */
  boolean enable_iwspword;
  /**
   * Dictionary entry to be added on "-iwspword" (-iwspentry) for N-gram
   */
  char *iwspentry;
  
#ifdef SEPARATE_BY_UNIGRAM
  /**
   * Number of best frequency words to be separated (linearized)
   * from lexicon tree (-sepnum)
   */
  int separate_wnum;
#endif

  /**
   * For isolated word recognition mode: name of head silence model
   */
  char wordrecog_head_silence_model_name[MAX_HMMNAME_LEN];
  /**
   * For isolated word recognition mode: name of tail silence model
   */
  char wordrecog_tail_silence_model_name[MAX_HMMNAME_LEN];
  /**
   * For isolated word recognition mode: name of silence as phone context
   */
  char wordrecog_silence_context_name[MAX_HMMNAME_LEN];

  /**
   * Name string of Unknown word for N-gram
   */
  char unknown_name[UNK_WORD_MAXLEN];

  /**
   * Pointer to next instance
   * 
   */
  struct __jconf_lm__ *next;
  
} JCONF_LM;

/**
 * Search parameters
 * 
 */
typedef struct __jconf_search__ {

  /**
   * Unique ID
   * 
   */
  short id;

  /**
   * Unique name
   * 
   */
  char name[JCONF_MODULENAME_MAXLEN];

  /**
   * Which AM configuration to refer
   * 
   */
  JCONF_AM *amconf;

  /**
   * Which LM configuration to refer
   * 
   */
  JCONF_LM *lmconf;

  /**
   * Compute only 1pass (-1pass)
   */
  boolean compute_only_1pass;
    
  /**
   * context handling
   */
  boolean ccd_handling;
    
  /**
   * force context-dependent handling
   */
  boolean force_ccd_handling;

  /**
   * LM weights
   * 
   */
  struct {
    /**
     * N-gram Language model weight (-lmp)
     */
    LOGPROB lm_weight;  
    /**
     * N-gram Word insertion penalty (-lmp)
     */
    LOGPROB lm_penalty; 
    /**
   * N-gram Language model weight for 2nd pass (-lmp2)
   */
    LOGPROB lm_weight2; 
    /**
     * N-gram Word insertion penalty for 2nd pass (-lmp2)
     */
    LOGPROB lm_penalty2;        
    /**
     * N-gram Additional insertion penalty for transparent words (-transp)
     */
    LOGPROB lm_penalty_trans;
    
    /**
     * Word insertion penalty for DFA grammar on 1st pass (-penalty1)
     */
    LOGPROB penalty1;
    /**
     * Word insertion penalty for DFA grammar on 2nd pass (-penalty2)
     */
    LOGPROB penalty2;
    
    /**
     * INTERNAL: TRUE if -lmp2 specified
     */
    boolean lmp2_specified;
    
    /**
     * INTERNAL: TRUE if -lmp specified
     */
    boolean lmp_specified;
  } lmp;
    
  /**
   * First pass parameters
   * 
   */
  struct {
    /**
     * Beam width of 1st pass. If value is -1 (not specified), system
     * will guess the value from dictionary size.  If 0, a possible
     * maximum value will be assigned to do full search.
     */
    int specified_trellis_beam_width;
    
#if defined(WPAIR) && defined(WPAIR_KEEP_NLIMIT)
    /**
     * Keeps only N token on word-pair approximation (-nlimit)
     */
    int wpair_keep_nlimit;
#endif

#ifdef HASH_CACHE_IW
    /**
     * Inter-word LM cache size rate (-iwcache)
     */
    int iw_cache_rate;
#endif

    /**
     * (DEBUG) use old build_wchmm() instead of build_wchmm2() for lexicon
     * construction (-oldtree)
     */
    boolean old_tree_function_flag;

#ifdef DETERMINE
    /**
     * (EXPERIMENTAL) score threshold between maximum node score and
     * maximum word end score for early word determination
     * 
     */
    LOGPROB determine_score_thres;

    /**
     * (EXPERIMENTAL) frame duration threshold for early word determination
     * 
     */
    int determine_duration_thres;

#endif /* DETERMINE */


  } pass1;

  /**
   * Second pass parameters
   * 
   */
  struct {
    /**
     * Search until N-best sentences are found (-n). Also see "-output".
     */
    int nbest;                
    /**
     * Word beam width of 2nd pass. -1 means no beaming (-b2)
     */
    int enveloped_bestfirst_width;
#ifdef SCAN_BEAM
    /**
     * Score beam threshold of 2nd pass (-sb)
     */
    LOGPROB scan_beam_thres;
#endif
    /**
     * Hypothesis overflow threshold at 2nd pass (-m)
     */
    int hypo_overflow;
    /**
     * Hypothesis stack size of 2nd pass (-s)
     */
    int stack_size;
    /**
     * Get next words from word trellis with a range of this frames
     * on 2nd pass (-lookuprange)
     */
    int lookup_range;
    
    /**
     * Limit expansion words for trellis words on neighbor frames
     * at 2nd pass of DFA for speedup (-looktrellis)
     */
    boolean looktrellis_flag;
    
  } pass2;

  /**
   * Word graph output
   * 
   */
  struct {

    /**
     * GraphOut: if enabled, graph search is enabled.
     * 
     */
    boolean enabled;

    /**
     * GraphOut: if enabled, output word graph
     * 
     */
    boolean lattice;

    /**
     * GraphOut: if enabled, generate confusion network
     * 
     */
    boolean confnet;

    /**
     * GraphOut: allowed margin for post-merging on word graph generation
     * (-graphrange) if set to -1, same word with different phone context
     * will be separated.
     */
    int graph_merge_neighbor_range;

#ifdef   GRAPHOUT_DEPTHCUT
    /**
     * GraphOut: density threshold to cut word graph at post-processing.
     * (-graphcut)  Setting larger value is safe for all condition.
     */
    int graphout_cut_depth;
#endif

#ifdef   GRAPHOUT_LIMIT_BOUNDARY_LOOP
    /**
     * GraphOut: limitation of iteration loop for word boundary adjustment
     * (-graphboundloop)
     */
    int graphout_limit_boundary_loop_num;
#endif

#ifdef GRAPHOUT_SEARCH_DELAY_TERMINATION
    /**
     * GraphOut: delay the termination of search on graph merging until
     * at least one sentence candidate is found
     * (-graphsearchdelay / -nographsearchdelay)
     */
    boolean graphout_search_delay;
#endif

  } graph;
  
  /**
   * Successive decoding (--enable-sp-segment)
   * 
   */
  struct {

    /**
     * TRUE if short-pause segmentation enabled for this instance
     * 
     */
    boolean enabled;