main_align.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 1,189 行 · 第 1/3 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.
 *
 * ====================================================================
 *
 */
/*
 * align-main.c -- Main driver routine for time alignment.
 * 
 * **********************************************
 * CMU ARPA Speech Project
 *
 * Copyright (c) 1996 Carnegie Mellon University.
 * ALL RIGHTS RESERVED.
 * **********************************************
 * 
 * HISTORY
 * 
 * 19-Jun-1998	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Modified to handle the new libfeat interface.
 * 
 * 11-Oct-1997	M K Ravishankar ([e-mail]) at Carnegie Mellon University.
 * 		Added MLLR transformation for each utterance.
 * 
 * 06-Mar-1997	M K Ravishankar ([e-mail]) at Carnegie Mellon University.
 * 		Added .semi. and .cont. options to -senmgaufn flag.
 *  
 * 16-Oct-1996	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added orig_stdout, orig_stderr hack to avoid hanging on exit under Linux.
 *  
 * 14-Oct-1996	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Removed explicit addition of SILENCE_WORD, START_WORD and
 * 		FINISH_WORD to the dictionary.
 * 
 * 18-Sep-1996	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]].
 * 		(There are incompatibilities with ,CTL output directory specification.)
 * 
 * 13-Sep-1996	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Normalized senone scores (subtracting the best) rather than density scores.
 * 		Bugfix: Absolute scores written to state score output file by removing
 * 		normalization factor.
 * 
 * 22-Jul-1996	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added absolute (unnormalized) acoustic scores in log file.
 * 		Added Sphinx-II compatible output segmentation files.
 * 
 * 15-Jul-1996	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Created.
 */


#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#include <s3types.h>

#include "kb.h"


#include "logs3.h"
#include "tmat.h"
#include "mdef.h"
#include "dict.h"
#include "agc.h"
#include "cmn.h"
#include "bio.h"
#include "feat.h"

/* ARCHAN: Dangerous routine :-)*/
#include "s3_align.h"
#include "ms_mllr.h"
#include "ms_gauden.h"
#include "ms_senone.h"
#include "interp.h"

/** \file main_align.c
   \brief Main driver routine for time alignment.
*/
static arg_t defn[] = {
    { "-logbase",
      ARG_FLOAT32,
      "1.0003",
      "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" },
    { "-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" },
    { "-log3table",
      ARG_INT32,
      "1",
      "Determines whether to use the log3 table or to compute the values at run time."},
    { "-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",
      "Variance normalize each utterance (yes/no; only applicable if CMN is also performed)" },
    { "-feat",	/* Captures the computation for converting input to feature vector */
      ARG_STRING,
      "1s_c_d_dd",
      "Feature stream: s2_4x / s3_1x39 / cep_dcep[,%d] / cep[,%d] / %d,%d,...,%d" },
    { "-dict",
      ARG_STRING,
      NULL,
      "Main pronunciation dictionary (lexicon) input file" },
    { "-fdict",
      ARG_STRING,
      NULL,
      "Optional filler word (noise word) pronunciation dictionary input file" },
    { "-compwd",
      ARG_INT32,
      "0",
      "Compound words in dictionary (not supported yet)" },
    { "-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,
      0,
      "No. of utterances in -ctl file to be processed (after -ctloffset).  Default: Until EOF" },
    { "-cepdir",
      ARG_STRING,
      ".",
      "Directory for utterances in -ctl file (if relative paths specified)." },
    { "-cepext",
      ARG_STRING,
      ".mfc",
      "File extension appended to utterances listed in ctl file" },

    { "-mllrctl",
      ARG_STRING,
      NULL,
      "Input control file listing MLLR input data; parallel to ctl argument file" },

    { "-lambda",
      ARG_STRING,
      NULL,
      "Interpolation weights (CD/CI senone) parameters input file" },

    { "-topn",
      ARG_INT32,
      "4",
      "No. of top scoring densities computed in each mixture gaussian codebook" },
    { "-beam",
      ARG_FLOAT64,
      "1e-64",
      "Main pruning beam applied to triphones in forward search" },
    { "-insent",
      ARG_STRING,
      NULL,
      "Input transcript file corresponding to control file" },
    { "-outsent",
      ARG_STRING,
      NULL,
      "Output transcript file with exact pronunciation/transcription" },
    { "-stsegdir",
      ARG_STRING,
      NULL,
      "Output directory for state segmentation files; optionally end with ,CTL" },
    { "-phsegdir",
      ARG_STRING,
      NULL,
      "Output directory for phone segmentation files; optionally end with ,CTL" },
    { "-wdsegdir",
      ARG_STRING,
      NULL,
      "Output directory for word segmentation files; optionally end with ,CTL" },
    { "-s2stsegdir",
      ARG_STRING,
      NULL,
      "Output directory for Sphinx-II format state segmentation files; optionally end with ,CTL" },
    { "-logfn",
      ARG_STRING,
      NULL,
      "Log file (default stdout/stderr)" },
    
    { NULL, ARG_INT32, NULL, NULL }
};

static mdef_t *mdef;		/* Model definition */
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 s3wid_t startwid, finishwid, silwid;

static int32 *senscale;		/* ALL senone scores scaled by senscale[i] in frame i */

static FILE *outsentfp = NULL;

/* For profiling/timing */
enum { tmr_utt, tmr_gauden, tmr_senone, tmr_align };
ptmr_t timers[5];

static int32 tot_nfr;
static ptmr_t tm_utt;

/*
 * Load and cross-check all models (acoustic/lexical/linguistic).
 */
static dict_t *dict;

static void models_init ( void )
{
    float32 varfloor, mixwfloor, tpfloor;
    int32 i;
    char *arg;
    
    /* 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"),
		      '_');	/* Compound word separator.  Default: none. */

    /* HACK!! Make sure SILENCE_WORD, START_WORD and 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!! */

    /* Codebooks */
    varfloor = *((float32 *) cmd_ln_access("-varfloor"));
    g = gauden_init ((char *) cmd_ln_access("-mean"),
		     (char *) cmd_ln_access("-var"),
		     varfloor);

    /* Verify codebook feature dimensions against libfeat */
    if (feat_n_stream(fcb) != g->n_feat) {
	E_FATAL("#feature mismatch: feat= %d, mean/var= %d\n",
		feat_n_stream(fcb), g->n_feat);
    }
    for (i = 0; i < feat_n_stream(fcb); i++) {
	if (feat_stream_len(fcb,i) != g->featlen[i]) {
	    E_FATAL("featlen[%d] mismatch: feat= %d, mean/var= %d\n", i,
		    feat_stream_len(fcb, i), g->featlen[i]);
	}
    }
    
    /* Senone mixture weights */
    mixwfloor = *((float32 *) cmd_ln_access("-mwfloor"));
    sen = senone_init ((char *) cmd_ln_access("-mixw"),
		       (char *) cmd_ln_access("-senmgau"),
		       mixwfloor);
    
    /* Verify senone parameters against gauden parameters */
    if (sen->n_feat != g->n_feat)
	E_FATAL("#Feature mismatch: gauden= %d, senone= %d\n", g->n_feat, sen->n_feat);
    if (sen->n_cw != g->n_density)
	E_FATAL("#Densities mismatch: gauden= %d, senone= %d\n", g->n_density, sen->n_cw);
    if (sen->n_gauden > g->n_mgau)
	E_FATAL("Senones need more codebooks (%d) than present (%d)\n",
		sen->n_gauden, g->n_mgau);
    if (sen->n_gauden < g->n_mgau)
	E_ERROR("Senones use fewer codebooks (%d) than present (%d)\n",
		sen->n_gauden, g->n_mgau);

    /* Verify senone parameters against model definition parameters */
    if (mdef->n_sen != sen->n_sen)
	E_FATAL("Model definition has %d senones; but #senone= %d\n",
		mdef->n_sen, sen->n_sen);

    /* CD/CI senone interpolation weights file, if present */
    if ((arg = (char *) cmd_ln_access ("-lambda")) != NULL) {
	interp = interp_init (arg);

	/* Verify interpolation weights size with senones */
	if (interp->n_sen != sen->n_sen)
	    E_FATAL("Interpolation file has %d weights; but #senone= %d\n",
		    interp->n_sen, sen->n_sen);
    } else
	interp = NULL;

    /* Transition matrices */
    tpfloor = *((float32 *) cmd_ln_access("-tpfloor"));
    tmat = tmat_init ((char *) cmd_ln_access("-tmat"), tpfloor);

    /* Verify transition matrices parameters against model definition parameters */
    if (mdef->n_tmat != tmat->n_tmat)
	E_FATAL("Model definition has %d tmat; but #tmat= %d\n",
		mdef->n_tmat, tmat->n_tmat);
    if (mdef->n_emit_state != tmat->n_state)
	E_FATAL("#Emitting states in model definition = %d, #states in tmat = %d\n",
		mdef->n_emit_state, tmat->n_state);

    arg = (char *) cmd_ln_access ("-agc");
    if ((strcmp (arg, "max") != 0) && (strcmp (arg, "none") != 0))
	E_FATAL("Unknown -agc argument: %s\n", arg);
    arg = (char *) cmd_ln_access ("-cmn");
    if ((strcmp (arg, "current") != 0) && (strcmp (arg, "none") != 0))
	E_FATAL("Unknown -cmn argument: %s\n", arg);
}


/*
 * Build a filename int buf as follows (without file extension):
 *     if dir ends with ,CTL and ctlspec does not begin with /, filename is dir/ctlspec
 *     if dir ends with ,CTL and ctlspec DOES begin with /, filename is ctlspec
 *     if dir does not end with ,CTL, filename is dir/uttid,
 * where ctlspec is the complete utterance spec in the input control file, and
 * uttid is the last component of ctlspec.
 */
static void build_output_uttfile (char *buf, char *dir, char *uttid, char *ctlspec)
{
    int32 k;
    
    k = strlen(dir);
    if ((k > 4) && (strcmp (dir+k-4, ",CTL") == 0)) {	/* HACK!! Hardwired ,CTL */
	if (ctlspec[0] != '/') {