julius/define.h

Go to the documentation of this file.

/*
 * Copyright (c) 1991-2006 Kawahara Lab., Kyoto University
 * Copyright (c) 2000-2005 Shikano Lab., Nara Institute of Science and Technology
 * Copyright (c) 2005-2006 Julius project team, Nagoya Institute of Technology
 * All rights reserved
 */

/*****************************************************************************/
/*****************************************************************************/

/* switch N-gram mode (julius) <-> grammar mode (julian) */
#ifdef USE_DFA
#undef  USE_NGRAM
#else
#define USE_NGRAM
#endif

/* delete incoherent option */
#ifdef USE_DFA
#ifdef UNIGRAM_FACTORING
#undef UNIGRAM_FACTORING
#endif
#define CATEGORY_TREE
#else  /* USE_NGRAM */
#ifdef CATEGORY_TREE
#undef CATEGORY_TREE
#endif
#endif /* USE_DFA */

/* abbreviations for verbose message output */
#define VERMES if (verbose_flag) j_printerr

/* define this to report memory usage on exit (Linux only) */
#undef REPORT_MEMORY_USAGE

#ifdef USE_NGRAM
/*** tree construction ***/
/* With 1-best approximation, Constructing a single tree from all words
   causes much error by factoring.  Listing each word flatly with no
   tree-organization will not cause this error, but the network becomes
   much larger and, especially, the inter-word LM handling becomes much more
   complex (O(n^2)).  The cost may be eased by LM caching, but it needs much
   memory. */
/* This is a trade-off of accuracy and cost */
#define SHORT_WORD_LEN 2
#ifdef LOWMEM
/* don't separate, construct a single tree from all words */
/* root nodes are about 50 in monophone, cache size will be 5MB on max */
#define NO_SEPARATE_SHORT_WORD
#else
#ifdef LOWMEM2
/* experimental: separate words frequently appears in corpus (1-gram) */
/* root nodes will be "-sepnum num" + 50, cache size will be 10MB or so */
#define NO_SEPARATE_SHORT_WORD
#define SEPARATE_BY_UNIGRAM
#define DEFAULT_SEPARATE_WNUM 150
#else
/* separate all short words (<= 2 phonemes) */
/* root nodes are about 1100 in 20k (proportional to vocabulary),
   cache size will be about 100MB on max */
#endif /* LOWMEM2 */
#endif /* LOWMEM */

/*#define HASH_CACHE_IW*/
/* "./configure --enable-lowmem" defines NO_SEPARATE_SHORT_WORD instead */

#endif /* USE_NGRAM */

#ifdef USE_NGRAM
/* default language model weight and insertion penalty for pass1 and pass2 */
/* these values come from the best parameters in IPA evaluation result */
#define DEFAULT_LM_WEIGHT_MONO_PASS1   5.0
#define DEFAULT_LM_PENALTY_MONO_PASS1 -1.0
#define DEFAULT_LM_WEIGHT_MONO_PASS2   6.0
#define DEFAULT_LM_PENALTY_MONO_PASS2  0.0
#ifdef PASS1_IWCD
#define DEFAULT_LM_WEIGHT_TRI_PASS1   8.0
#define DEFAULT_LM_PENALTY_TRI_PASS1 -2.0
#define DEFAULT_LM_WEIGHT_TRI_PASS2   8.0
#define DEFAULT_LM_PENALTY_TRI_PASS2 -2.0
#else
#define DEFAULT_LM_WEIGHT_TRI_PASS1   9.0
#define DEFAULT_LM_PENALTY_TRI_PASS1  8.0
#define DEFAULT_LM_WEIGHT_TRI_PASS2  11.0
#define DEFAULT_LM_PENALTY_TRI_PASS2 -2.0
#endif /* PASS1_IWCD */
#endif /* USE_NGRAM */

/* Switch head/tail word insertion penalty to be inserted */
#undef FIX_PENALTY

/* some definitions for short-pause segmentation */
#ifdef SP_BREAK_CURRENT_FRAME
#undef SP_BREAK_EVAL            /* output messages for evaluation */
#undef SP_BREAK_DEBUG           /* output messages for debug */
#undef SP_BREAK_RESUME_WORD_BEGIN /* resume word = maxword at beginning of sp area */
#endif

/* '01/10/18 by ri: enable fix for trellis lookup order */
#define PREFER_CENTER_ON_TRELLIS_LOOKUP

#ifdef MULTIPATH_VERSION
/* '01/11/28 by ri: malloc step for startnode */
# define STARTNODE_STEP 300
/* default value of iwsp penalty */
# define IWSP_PENALTY_DEFAULT -1.0
#endif

/* default dict entry for IW-sp word that will be added to dict with -iwspword */
#ifdef USE_NGRAM
#define IWSPENTRY_DEFAULT "<UNK> [sp] sp sp"
#endif

/* confidence scoring method */
#ifdef CONFIDENCE_MEASURE
# ifndef CM_NBEST       /* use conventional N-best CM, will be defined if "--enable-cm-nbest" specified */
#  define CM_SEARCH     /* otherwise, use on-the-fly CM scoring */
# endif
#endif

/* dynamic word graph generation */
#ifdef GRAPHOUT        /* output result in word-graph format */

#undef GRAPHOUT_SEARCH_CONSIDER_RIGHT /* if defined, only hypothesis whose
                                         left/right contexts is already
                                         included in popped hypo will be merged.
                                         EXPERIMENTAL, should not be defined.
                                       */
#ifdef CM_SEARCH_LIMIT
#undef CM_SEARCH_LIMIT_AFTER    /* enable above only after 1 sentence found */
#undef CM_SEARCH_LIMIT_POP      /* terminate hypo of low CM on pop */
#endif

/* compute exact boundary instead of using 1st pass result */
/* also propagate exact time boundary to the right context after generation */
/* this may produce precise word boundary, but cause bigger word graph output */
#define GRAPHOUT_PRECISE_BOUNDARY

#undef GDEBUG                   /* enable debug message in graphout.c */

#endif /* GRAPHOUT */

/* some decoding fix candidates */
#undef FIX_35_PASS2_STRICT_SCORE /* fix hypothesis scores by enabling
                                      bt_discount_pescore() in standard mode
                                      with PASS2_STRICT_IWCD, 
                                   */
#define FIX_35_INHIBIT_SAME_WORD_EXPANSION /* privent connecting the same trellis word in 2nd pass */


/* below are new since 3.5.2 */
#ifdef GRAPHOUT

#define GRAPHOUT_OVERWRITE

/* with GRAPHOUT_OVERWRITE, use gscore_head instead of fscore_head */
#undef GRAPHOUT_OVERWRITE_GSCORE

#define GRAPHOUT_LIMIT_BOUNDARY_LOOP

#define GRAPHOUT_LIMIT_BOUNDARY_LOOP_NUM_DEFAULT 20

#define GRAPHOUT_SEARCH_DELAY_TERMINATION

#define GRAPHOUT_DEPTHCUT

#define GRAPHOUT_DEPTHCUT_DEFAULT 80

#endif /* GRAPHOUT */

---