main_decode_anytopo.c

来自「CMU大名鼎鼎的SPHINX－3大词汇量连续语音识别系统」· C语言 代码 · 共 1,466 行 · 第 1/4 页

/* ====================================================================
 * 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.
 *
 * ====================================================================
 *
 */
/*
 * main.c -- Main S3 decoder driver.
 *
 * **********************************************
 * CMU ARPA Speech Project
 *
 * Copyright (c) 1996 Carnegie Mellon University.
 * ALL RIGHTS RESERVED.
 * **********************************************
 * 
 * HISTORY
 * 
 * 26-Jul-04    ARCHAN ([e-mail]) at Carnegie Mellon Unversity
 *              First incorporated from sphinx 3.0 code base to 3.X codebase. 
 *
 * $Log: main_decode_anytopo.c,v $
 * Revision 1.7  2004/12/23 21:05:22  arthchan2003
 * Enable compilation of decode_anytopo, change option names from -match to -hyp, it makes the code more consistent.
 *
 * Revision 1.6  2004/12/14 00:50:33  arthchan2003
 * 1, Change the code to accept extension, 2, add timer to livepretend, 3, fixing the s3_astar to separate the bypass variable to bypass and is_filler_bypass.  4, Add some doxygen comments. 5, Don't care about changes in main_decode_anytopo.c. It is still under work, 6, remove option -help and -example from 3.5 releases.
 *
 * Revision 1.5  2004/12/06 11:31:47  arthchan2003
 * Fix brief comments for programs.
 *
 * Revision 1.4  2004/12/06 11:15:11  arthchan2003
 * Enable doxygen in the program directory.
 *
 * Revision 1.3  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 1.2  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 1.1  2004/11/14 07:00:08  arthchan2003
 * 1, Finally, a version of working flat decoder is completed. It is not compiled in the standard compilation yet because there are two many warnings. 2, eliminate the statics variables in  fe_sigproc.c
 *
 * Revision 1.2  2002/12/03 23:02:40  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.
 * 
 * 06-Mar-97	M K Ravishankar ([e-mail]) at Carnegie Mellon University.
 * 		Added .semi. and .cont. options to -senmgaufn flag.
 * 
 * 02-Dec-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Restricted MLLR transformation to CD mixture densities only.
 * 
 * 15-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed the meaning of -matchsegfn and, correspondingly, log_hypseg().
 * 
 * 11-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added -min_endfr and -dagfudge arguments.
 * 
 * 08-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added BSTXCT: reporting since that became available from dag_search.
 * 
 * 07-Nov-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added ,NODES suffix to -outlatdir argument for dumping only words to
 * 		lattice output files.
 *  
 * 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.
 * 
 * 04-Oct-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added unlimit() call to remove malloc restrictions.
 *  
 * 26-Sep-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added separate language weight (-bestpathlw) for bestpath DAG search.
 * 		Added -mllrctlfn flag and handling.
 * 
 * 21-Sep-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added -bptblsize argument.
 * 
 * 18-Sep-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added optional start/end frame specification in control file, for
 * 		processing selected segments (utterances) from a large cepfile.
 * 		Control spec: cepfile [startframe endframe [uttid]].
 * 
 * 13-Sep-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Bugfix: added senscale to bestscr before writing best score file.
 * 		(Otherwise, the scaled scores are meaningless.)
 * 		Added ,EXACT suffix option to -matchfn argument, and correspondingly
 * 		added "exact" argument to log_hypstr().  (But running bestpath search
 * 		will still cause <sil> and filler words to be removed from matchfile.)
 * 
 * 12-Sep-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed structure of gauden/senone computation:
 * 		    from: foreach (gauden) {eval gauden}; foreach (senone) {senone}
 * 		    to:   foreach (gauden) {eval gauden;  foreach (senone in gauden) {...}}
 * 		reducing memory space for results of gauden, specially in block mode.
 * 		Normalized senone scores (subtracting the best) rather than density scores.
 * 		Changed active senone list to flags.
 * 
 * 09-Sep-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Block-mode gauden computation for improving cache performance: changed
 *		    from: foreach(frame) {foreach(gauden)...}
 * 		    to:   foreach(gauden) {foreach(frame)...}
 * 		within a block of frames.  Must evaluate all gauden, not just active ones.
 * 		But even so the resulting caching performance is better.
 * 
 * 29-Aug-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed argument -inlatext to -latext.
 * 		Added check to ensure input and output lattice directories are different.
 * 		Added reporting of hostname.
 * 
 * 23-Aug-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed profiling to use timing_ functions, available on all platforms.
 * 		Added write_bestscore() for writing best statescore in each frame,
 * 		for helping determine desirable beamwidth.
 * 
 * 24-Jun-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added DAG search option and consolidated logging and reporting.
 * 		Added backtrace option.
 * 
 * 22-Jul-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added absolute (unnormalized) acoustic scores in log file.
 * 		Added uttid in log file with each word segmentation.
 * 		Compute only active codebooks and senones if multiple codebooks present.
 * 
 * 20-Jan-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added specification of input word lattices to limit search (-inlatdir
 * 		argument...).  Added computation of active senone and gauden codebook
 * 		lists when such a lattice is provided, to minimize computation.
 * 		Added -cmn and -agc flags.
 * 
 * 10-Jan-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added call to feat_init(); added cepsize variable and initialization;
 * 		Changed argument to norm_mean from featlen[0]+1 to cepsize.
 * 
 * 13-Dec-95	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed call to senone_eval to senone_eval_all optimized for the
 * 		semi-continuous case.
 * 		Completed handling multiple mixture-gaussian codebooks.
 * 
 * 01-Dec-95	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Created.
 */

/** \file main_decode_anytopo.c
 * \brief Main driver for sphinx 3.0 decoding (or the slow decoder)
 */

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#if (! WIN32)
#include <unistd.h>
#endif
#include <assert.h>

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

#include "cmn.h"
#include "agc.h"

#include "flat_fwd.h"
#include "ms_mllr.h"
#include "ms_gauden.h"
#include "ms_senone.h"
#include "interp.h"
#include "s3_dag.h"


static gauden_t *g;		/* Gaussian density codebooks */
static senone_t *sen;		/* Senones */
static interp_t *interp;	/* CD/CI interpolation */
static tmat_t *tmat;		/* HMM transition matrices */

static feat_t *fcb;           /* Feature type descriptor (Feature Control Block) */
static float32 ***feat = NULL;        /* Speech feature data */


static mdef_t *mdef;		/* Model definition */

extern lm_t* lm;
extern dict_t* dict;
extern fillpen_t* fpen;
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 int32 *senscale;		/* ALL senone scores scaled by senscale[i] in frame i */
static int32 *bestscr;		/* Best statescore in each frame */

ptmr_t tmr_utt;
ptmr_t tmr_fwdvit;
ptmr_t tmr_bstpth;
ptmr_t tmr_gausen;
ptmr_t tmr_fwdsrch;

pctr_t ctr_nfrm;
pctr_t ctr_nsen;


static int32 tot_nfr;

static char *inlatdir;
static char *outlatdir;
static int32 outlat_onlynodes;
static FILE *matchfp, *matchsegfp;
static int32 matchexact;


/*
 * Command line arguments.
 */
static arg_t defn[] = {
    { "-log3table",
      ARG_INT32,
      "1",
      "Determines whether to use the log3 table or to compute the values at run time."},
    { "-logbase",
      ARG_FLOAT32,
      "1.0001",
      "Base in which all log values calculated" },
    { "-mdef", 
      ARG_STRING,
      NULL,
      "Model definition input file: triphone -> senones/tmat tying" },
    { "-tmat",
      ARG_STRING,
      NULL,
      "Transition matrix input file" },
    { "-mean",
      ARG_STRING,
      NULL,
      "Mixture gaussian codebooks mean parameters input file" },
    { "-var",
      ARG_STRING,
      NULL,
      "Mixture gaussian codebooks variance parameters input file" },
    { "-senmgau",
      ARG_STRING,
      ".cont.",
      "Senone to mixture-gaussian mapping file (or .semi. or .cont.)" },
    { "-mixw",
      ARG_STRING,
      NULL,
      "Senone mixture weights parameters input file" },
    { "-lambda",
      ARG_STRING,
      NULL,
      "Interpolation weights (CD/CI senone) parameters input file" },
    { "-tpfloor",
      ARG_FLOAT32,
      "0.0001",
      "Triphone state transition probability floor applied to -tmat file" },
    { "-varfloor",
      ARG_FLOAT32,
      "0.0001",
      "Codebook variance floor applied to -var file" },
    { "-mwfloor",
      ARG_FLOAT32,
      "0.0000001",
      "Codebook mixture weight floor applied to -mixw file" },
    { "-agc",
      ARG_STRING,
      "max",
      "AGC.  max: C0 -= max(C0) in current utt; none: no AGC" },
    { "-cmn",
      ARG_STRING,
      "current",
      "Cepstral mean norm.  current: C[1..n-1] -= mean(C[1..n-1]) in current utt; none: no CMN" },
    { "-varnorm",
      ARG_STRING,
      "no",
      "Cepstral var norm. yes: C[0..n-1] /= stddev(C[0..n-1]); no = no norm */"},
    { "-feat",
      ARG_STRING,
      "s2_4x",
      "Feature stream:\n\t\t\t\ts2_4x: Sphinx-II type 4 streams, 12cep, 24dcep, 3pow, 12ddcep\n\t\t\t\ts3_1x39: Single stream, 12cep+12dcep+3pow+12ddcep\n\t\t\t\t1s_12c_12d_3p_12dd: Single stream, 12cep+12dcep+3pow+12ddcep\n\t\t\t\t1s_c: Single stream, given input vector only\n\t\t\t\t1s_c_d: Feature + Deltas only\n\t\t\t\t1s_c_dd: Feature + Double deltas only\n\t\t\t\t1s_c_d_dd: Feature + Deltas + Double deltas\n\t\t\t\t1s_c_wd_dd: Feature cep+windowed delcep+deldel \n\t\t\t1s_c_d_ld_dd: Feature + delta + longter delta + doubledelta" },
/* ADDED BY BHIKSHA: 6 JAN 98 */
    { "-lminmemory",
      ARG_INT32,
      "0",
      "Load language model into memory (default: use disk cache for lm"},
    { "-ceplen",
      ARG_INT32,
      "13",
      "Length of input feature vector" },
    { "-dict",
      ARG_STRING,
      NULL,
      "Main pronunciation dictionary (lexicon) input file" },
    { "-fdict",
      ARG_STRING,
      NULL,
      "Silence and filler (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" },