main_dag.c

CMU大名鼎鼎的SPHINX－3大词汇量连续语音识别系统

/* ====================================================================
 * Copyright (c) 1995-2004 Carnegie Mellon University.  All rights
 * reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer. 
 *
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in
 *    the documentation and/or other materials provided with the
 *    distribution.
 *
 * This work was supported in part by funding from the Defense Advanced 
 * Research Projects Agency and the National Science Foundation of the 
 * United States of America, and the CMU Sphinx Speech Consortium.
 *
 * THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY ``AS IS'' AND 
 * ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, 
 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY
 * NOR ITS EMPLOYEES BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 
 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * ====================================================================
 *
 */
/*
 * dag-main.c -- Driver for DAG search.
 *
 * **********************************************
 * CMU ARPA Speech Project
 *
 * Copyright (c) 1996 Carnegie Mellon University.
 * ALL RIGHTS RESERVED.
 * **********************************************
 * 
 * HISTORY
 * 
 * 
 * 
 * 27-Jul-04    ARCHAN ([e-mail]) at Carnegie Mellon Unversity
 *              First incorporate it from s3 code base
 *
 * 
 * $Log: main_dag.c,v $
 * Revision 1.8  2004/12/06 11:31:47  arthchan2003
 * Fix brief comments for programs.
 *
 * Revision 1.7  2004/12/06 11:15:11  arthchan2003
 * Enable doxygen in the program directory.
 *
 * Revision 1.6  2004/12/05 12:01:32  arthchan2003
 * 1, move libutil/libutil.h to s3types.h, seems to me not very nice to have it in every files. 2, Remove warning messages of main_align.c 3, Remove warning messages in chgCase.c
 *
 * Revision 1.5  2004/11/16 05:13:19  arthchan2003
 * 1, s3cipid_t is upgraded to int16 because we need that, I already check that there are no magic code using 8-bit s3cipid_t
 * 2, Refactor the ep code and put a lot of stuffs into fe.c (should be renamed to something else.
 * 3, Check-in codes of wave2feat and cepview. (cepview will not dump core but Evandro will kill me)
 * 4, Make the same command line frontends for decode, align, dag, astar, allphone, decode_anytopo and ep . Allow the use a file to configure the application.
 * 5, Make changes in test such that test-allphone becomes a repeatability test.
 * 6, cepview, wave2feat and decode_anytopo will not be installed in 3.5 RCIII
 * (Known bugs after this commit)
 * 1, decode_anytopo has strange bugs in some situations that it cannot find the end of the lattice. This is urgent.
 * 2, default argument file's mechanism is not yet supported, we need to fix it.
 * 3, the bug discovered by SonicFoundry is still not fixed.
 *
 * Revision 1.4  2004/10/07 22:46:26  dhdfu
 * Fix compiler warnings that are also real bugs (but why does this
 * function take an int32 when -lw is a float parameter?)
 *
 * Revision 1.3  2004/09/13 08:13:28  arthchan2003
 * update copyright notice from 200x to 2004
 *
 * Revision 1.2  2004/09/09 20:29:08  arthchan2003
 * Added test for astar and dag.  Put a hack in s3_dag.c that allows 0 as acoustice score.
 *
 * Revision 1.1  2004/08/20 08:25:19  arthchan2003
 * Sorry, I forget to add the main of dag.c
 *
 * Revision 1.2  2002/12/03 23:02:37  egouvea
 * Updated slow decoder with current working version.
 * Added copyright notice to Makefiles, *.c and *.h files.
 * Updated some of the documentation.
 *
 * Revision 1.1.1.1  2002/12/03 20:20:46  robust
 * Import of s3decode.
 *
 *
 * 08-Sep-97	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added .Z compression option to lattice files.
 * 
 * 22-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added -maxedge argument to control memory usage.
 * 
 * 21-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added -maxlmop and -maxlpf options to control execution time.
 * 
 * 15-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed the meaning of -matchsegfn and, correspondingly, log_hypseg().
 * 
 * 08-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added BSTXCT: reporting since that became available from dag_search.
 * 
 * 05-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added -dagfudge and -min_endfr parameter handling.
 *  
 * 16-Oct-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added orig_stdout, orig_stderr hack to avoid hanging on exit under Linux.
 *  
 * 11-Oct-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added fillpen_init() and removed explicit addition of SILENCE_WORD,
 * 		START_WORD and FINISH_WORD to the dictionary.
 * 
 * 03-Sep-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Creating.
 */

/** \file main_dag.c
 * \brief main driver for DAG and find the best path. 
 */
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#if (! WIN32)
#include <unistd.h>
#endif
#include <assert.h>

#include <s3types.h>
#include "s3_dag.h"
#include "logs3.h"
#include "tmat.h"
#include "mdef.h"
#include "dict.h"
#include "lm.h"
#include "fillpen.h"
#include "search.h"
#include "wid.h"


static mdef_t *mdef;		/* Model definition */
extern dict_t *dict;            /* The dictionary */
extern lm_t *lm ;               /* The lm */
extern fillpen_t *fpen;         /* The filler penalty structure */

extern s3lmwid_t *dict2lmwid;   /* Mapping from decoding dictionary wid's to lm 
ones.  They may not be the same! */

static s3wid_t startwid, finishwid, silwid;

static ptmr_t tm_utt;
static int32 tot_nfr;


/*
 * Command line arguments.
 */
static arg_t defn[] = {
    { "-logbase",
      ARG_FLOAT32,
      "1.0001",
      "Base in which all log values calculated" },
    { "-lminmemory",
      ARG_INT32,
      "0",
      "Load language model into memory (default: use disk cache for lm"},
    { "-log3table",
      ARG_INT32,
      "1",
      "Determines whether to use the log3 table or to compute the values at run time."},
    { "-mdef",
      ARG_STRING,
      NULL,
      "Model definition input file: triphone -> senones/tmat tying" },
    { "-dict",
      ARG_STRING,
      NULL,
      "Main pronunciation dictionary (lexicon) input file" },
    { "-fdict",
      ARG_STRING,
      NULL,
      "Optional filler word (noise word) pronunciation dictionary input file" },
    { "-lm",
      ARG_STRING,
      NULL,
      "Language model input file (precompiled .DMP file)" },
    { "-lw",
      ARG_FLOAT32,
      "9.5",
      "Language weight: empirical exponent applied to LM probabilty" },
    { "-ugwt",
      ARG_FLOAT32,
      "0.7",
      "LM unigram weight: unigram probs interpolated with uniform distribution with this weight" },
    { "-inspen",
      ARG_FLOAT32,
      "0.65",
      "Word insertion penalty" },
    { "-ctl",
      ARG_STRING,
      NULL,
      "Input control file listing utterances to be decoded" },
    { "-ctloffset",
      ARG_INT32,
      "0",
      "No. of utterances at the beginning of -ctl file to be skipped" },
    { "-ctlcount",
      ARG_INT32,
      NULL,
      "No. of utterances in -ctl file to be processed (after -ctloffset).  Default: Until EOF" },
    { "-silpen",
      ARG_FLOAT32,
      "0.1",
      "Language model 'probability' of silence word" },
    { "-noisepen",
      ARG_FLOAT32,
      "0.05",
      "Language model 'probability' of each non-silence filler word" },
    { "-fillpen",
      ARG_STRING,
      NULL,
      "Filler word probabilities input file (used in place of -silpen and -noisepen)" },
    { "-min_endfr",
      ARG_INT32,
      "3",
      "Nodes ignored during search if they persist for fewer than so many end frames" },
    { "-dagfudge",
      ARG_INT32,
      "2",
      "(0..2); 1 or 2: add edge if endframe == startframe; 2: if start == end-1" },
    { "-maxlpf",
      ARG_INT32,
      "40000",
      "Max LMops/frame after which utterance aborted; controls CPU use (see maxlmop)" },
    { "-maxlmop",
      ARG_INT32,
      "100000000",
      "Max LMops in utterance after which it is aborted; controls CPU use (see maxlpf)" },
    { "-maxedge",
      ARG_INT32,
      "2000000",
      "Max DAG edges allowed in utterance; aborted if exceeded; controls memory usage" },
    { "-inlatdir",
      ARG_STRING,
      NULL,
      "Input word-lattice directory with per-utt files for restricting words searched" },
    { "-latext",
      ARG_STRING,
      "lat.gz",
      "Word-lattice filename extension (.gz or .Z extension for compression)" },
    { "-match",
      ARG_STRING,
      NULL,
      "Recognition result output file (old NIST format; pre Nov95)" },
    { "-matchseg",
      ARG_STRING,
      NULL,
      "Exact recognition result file with word segmentations and scores" },
    { "-logfn",
      ARG_STRING,
      NULL,
      "Log file (default stdout/stderr)" },
    { "-backtrace",
      ARG_INT32,
      "1",
      "Whether detailed backtrace information (word segmentation/scores) shown in log" },
    
    { NULL, ARG_INT32,  NULL, NULL }
};

/*
 * Load and cross-check all models (acoustic/lexical/linguistic).
 */
static void models_init ( void )
{
    /* HMM model definition */
    mdef = mdef_init ((char *) cmd_ln_access("-mdef"));

    /* Dictionary */
    dict = dict_init (mdef,
		      (char *) cmd_ln_access("-dict"),
		      (char *) cmd_ln_access("-fdict"),
		      0);

    /* HACK!! Make sure S3_SILENCE_WORD, S3_START_WORD and S3_FINISH_WORD are in dictionary */
    silwid = dict_wordid (dict,S3_SILENCE_WORD);
    startwid = dict_wordid (dict, S3_START_WORD);
    finishwid = dict_wordid (dict, S3_FINISH_WORD);
    if (NOT_S3WID(silwid) || NOT_S3WID(startwid) || NOT_S3WID(finishwid)) {
	E_FATAL("%s, %s, or %s missing from dictionary\n",
		S3_SILENCE_WORD, S3_START_WORD, S3_FINISH_WORD);
    }
    if ((dict->filler_start > dict->filler_end) || (! dict_filler_word (dict,silwid)))
	E_FATAL("%s must occur (only) in filler dictionary\n", S3_SILENCE_WORD);
    /* No check that alternative pronunciations for filler words are in filler range!! */

    /* LM */

    {
      char *lmfile;

      lmfile = (char *) cmd_ln_access("-lm");
      if (! lmfile)
	E_FATAL("-lm argument missing\n");

      lm = lm_read (lmfile, 
		    *(float32 *)cmd_ln_access("-lw"),
		    *(float32 *)cmd_ln_access("-inspen"),
		    *(float32 *)cmd_ln_access("-ugwt"));



      /* Filler penalties */
      fpen = fillpen_init (dict,(char *) cmd_ln_access("-fillpen"),
		    *(float32 *)cmd_ln_access("-silpen"),
		    *(float32 *)cmd_ln_access("-noisepen"),
		    *(float32 *)cmd_ln_access("-lw"),
		    *(float32 *)cmd_ln_access("-inspen"));
    }

    dict2lmwid = wid_dict_lm_map(dict, lm, *(float32 *)cmd_ln_access("-lw"));
}