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

	 */
	/* Obtain active senone flags */
	ptmr_start (timers+tmr_senone);
	align_sen_active (sen_active, sen->n_sen);
	/* Turn active flags into list (for faster access) */

	if (interp) {
	    for (s = 0; s < mdef->n_ci_sen; s++)
		sen_active[s] = 1;
	}

	n_sen_active = 0;
	for (s = 0; s < mdef->n_sen; s++) {
	    if (sen_active[s])
		sen_active[n_sen_active++] = s;
	}
	ptmr_stop (timers+tmr_senone);
	
	/* Flag all active mixture-gaussian codebooks */
	ptmr_start (timers+tmr_gauden);
	for (gid = 0; gid < g->n_mgau; gid++)
	    mgau_active[gid] = 0;
	for (s = 0; s < n_sen_active; s++) {
	    sid = sen_active[s];
	    mgau_active[sen->mgau[sid]] = 1;
	}
	
	/* Compute topn gaussian density values (for active codebooks) */
	for (gid = 0; gid < g->n_mgau; gid++)
	    if (mgau_active[gid])
		gauden_dist (g, gid, topn, fv, dist[gid]);
	ptmr_start (timers+tmr_gauden);
	
	/* Evaluate active senones */
	ptmr_start (timers+tmr_senone);
	best = (int32) 0x80000000;
	for (s = 0; s < n_sen_active; s++) {
	    sid = sen_active[s];
	    senscr[sid] = senone_eval (sen, sid, dist[sen->mgau[sid]], topn);
	    if (best < senscr[sid])
		best = senscr[sid];
	}

	if (interp) {
	    for (s = 0; s < n_sen_active; s++) {
		if ((sid = sen_active[s]) >= mdef->n_ci_sen)
		    interp_cd_ci (interp, senscr, sid, mdef->cd2cisen[sid]);
	    }
	}


	/* Normalize senone scores (interpolation above can only lower best score) */
	for (s = 0; s < n_sen_active; s++) {
	    sid = sen_active[s];
	    senscr[sid] -= best;
	}
	senscale[i] = best;
	ptmr_stop (timers+tmr_senone);
	
	/* Step alignment one frame forward */
	ptmr_start (timers+tmr_align);
	align_frame (senscr);
	ptmr_stop (timers+tmr_align);
	ptmr_stop (timers+tmr_utt);
    }
    ptmr_stop (&tm_utt);

    printf ("\n");

    /* Wind up alignment for this utterance */
    if (align_end_utt (&stseg, &phseg, &wdseg) < 0)
	E_ERROR("Final state not reached; no alignment for %s\n\n", uttid);
    else {
	if (s2stsegdir)
	    write_s2stseg (s2stsegdir, stseg, uttid, ctlspec);
	if (stsegdir)
	    write_stseg (stsegdir, stseg, uttid, ctlspec);
	if (phsegdir)
	    write_phseg (phsegdir, phseg, uttid, ctlspec);
	if (wdsegdir)
	    write_wdseg (wdsegdir, wdseg, uttid, ctlspec);
	if (outsentfp)
	    write_outsent (outsentfp, wdseg, uttid);
    }
    
    align_destroy_sent_hmm ();
    
    ptmr_print_all (stdout, timers, nfr*0.1);

    printf("EXECTIME: %5d frames, %7.2f sec CPU, %6.2f xRT; %7.2f sec elapsed, %6.2f xRT\n",
	   nfr,
	   tm_utt.t_cpu, tm_utt.t_cpu * 100.0 / nfr,
	   tm_utt.t_elapsed, tm_utt.t_elapsed * 100.0 / nfr);

    tot_nfr += nfr;
}


#define UPPER_CASE(c)   ((((c) >= 'a') && ((c) <= 'z')) ? (c-32) : c)
/* Case insensitive string compare */
static int32 id_cmp (char *str1, char *str2)
{
    char c1, c2;
    
    for (;;) {
        c1 = *(str1++);
        c1 = UPPER_CASE(c1);
        c2 = *(str2++);
        c2 = UPPER_CASE(c2);
        if (c1 != c2)
            return (c1-c2);
        if (c1 == '\0')
            return 0;
    }
}


/* Process utterances in the control file (ctl argument) */
static void process_ctlfile ( void )
{
    FILE *ctlfp, *sentfp, *mllrctlfp;
    char *ctlfile, *cepdir, *cepext, *sentfile, *outsentfile, *mllrctlfile;
    char line[1024], ctlspec[1024];
    int32 ctloffset, ctlcount, sf, ef, nfr;
    char mllrfile[4096], prevmllr[4096], sent[16384];
    char uttid[1024];
    int32 i, k;
    
    ctlfile = (char *) cmd_ln_access("-ctl");
    if ((ctlfp = fopen (ctlfile, "r")) == NULL)
	E_FATAL("fopen(%s,r) failed\n", ctlfile);
    
    if ((mllrctlfile = (char *) cmd_ln_access("-mllrctl")) != NULL) {
	if ((mllrctlfp = fopen (mllrctlfile, "r")) == NULL)
	    E_FATAL("fopen(%s,r) failed\n", mllrctlfile);
    } else
	mllrctlfp = NULL;
    prevmllr[0] = '\0';
    
    sentfile = (char *) cmd_ln_access("-insent");
    if ((sentfp = fopen (sentfile, "r")) == NULL)
	E_FATAL("fopen(%s,r) failed\n", sentfile);

    if ((outsentfile = (char *) cmd_ln_access("-outsent")) != NULL) {
	if ((outsentfp = fopen (outsentfile, "w")) == NULL)
	    E_FATAL("fopen(%s,r) failed\n", outsentfile);
    }
    
    E_INFO("Processing ctl file %s\n", ctlfile);
    
    cepdir = (char *) cmd_ln_access("-cepdir");
    cepext = (char *) cmd_ln_access("-cepext");
    assert ((cepdir != NULL) && (cepext != NULL));
    
    ctloffset = *((int32 *) cmd_ln_access("-ctloffset"));
    if (! cmd_ln_access("-ctlcount"))
	ctlcount = 0x7fffffff;	/* All entries processed if no count specified */
    else
	ctlcount = *((int32 *) cmd_ln_access("-ctlcount"));
    if (ctlcount == 0) {
	E_INFO("-ctlcount argument = 0!!\n");
	fclose (ctlfp);
	fclose (sentfp);
	if (outsentfp)
	    fclose (outsentfp);
	
	return;
    }
    
    /* Skipping initial offset */
    if (ctloffset > 0)
	E_INFO("Skipping %d utterances in the beginning of control file\n",
	       ctloffset);
    while ((ctloffset > 0) && (fgets(line, sizeof(line), ctlfp) != NULL)) {
	if (sscanf (line, "%s", ctlspec) > 0) {
	    if (fgets (sent, sizeof(sent), sentfp) == NULL) {
		E_ERROR("EOF(%s)\n", sentfile);
		ctlcount = 0;
		break;
	    }
	    if (mllrctlfp) {
		if (fscanf (mllrctlfp, "%s", mllrfile) != 1)
		    E_FATAL ("Unexpected EOF(%s)\n", mllrctlfile);
	    }
	    --ctloffset;
	}
    }

    /* Process the specified number of utterance or until end of control file */
    while ((ctlcount > 0) && (fgets(line, sizeof(line), ctlfp) != NULL)) {
	printf ("\n");
	E_INFO("Utterance: %s", line);
	
	sf = 0;
	ef = (int32)0x7ffffff0;
	if ((k = sscanf (line, "%s %d %d %s", ctlspec, &sf, &ef, uttid)) <= 0)
	    continue;	    /* Empty line */

	if ((k == 2) || ( (k >= 3) && ((sf >= ef) || (sf < 0))) )
	    E_FATAL("Bad ctlfile line: %s\n", line);

	if (k < 4) {
	    /* Create utt-id from mfc-filename (and sf/ef if specified) */
	    for (i = strlen(ctlspec)-1; (i >= 0) && (ctlspec[i] != '/'); --i);
	    if (k == 3)
		sprintf (uttid, "%s_%d_%d", ctlspec+i+1, sf, ef);
	    else
		strcpy (uttid, ctlspec+i+1);
	}

	if (mllrctlfp) {
	    if (fscanf (mllrctlfp, "%s", mllrfile) != 1)
		E_FATAL ("Unexpected EOF(%s)\n", mllrctlfile);
	    
	    if (strcmp (prevmllr, mllrfile) != 0) {
		float32 ***A, **B;
		int32 gid, sid;
		uint8 *mgau_xform;
		
		gauden_mean_reload (g, (char *) cmd_ln_access("-mean"));
		
		if (ms_mllr_read_regmat (mllrfile, &A, &B,
				      fcb->stream_len, feat_n_stream(fcb)) < 0)
		    E_FATAL("ms_mllr_read_regmat failed\n");
		
		mgau_xform = (uint8 *) ckd_calloc (g->n_mgau, sizeof(uint8));

		/* Transform each non-CI mixture Gaussian */
		for (sid = 0; sid < sen->n_sen; sid++) {
		    if (mdef->cd2cisen[sid] != sid) {	/* Otherwise it's a CI senone */
			gid = sen->mgau[sid];
			if (! mgau_xform[gid]) {
			    ms_mllr_norm_mgau (g->mean[gid], g->n_density, A, B,
					    fcb->stream_len, feat_n_stream(fcb));
			    mgau_xform[gid] = 1;
			}
		    }
		}

		ckd_free (mgau_xform);
		
		ms_mllr_free_regmat (A, B, fcb->stream_len, feat_n_stream(fcb));

		strcpy (prevmllr, mllrfile);
	    }
	}


	/* Read utterance transcript */
	if (fgets (sent, sizeof(sent), sentfp) == NULL) {
	    E_ERROR("EOF(%s)\n", sentfile);
	    break;
	}
	/* Strip utterance id from the end of the transcript */
	for (k = strlen(sent) - 1;
	     (k > 0) && ((sent[k] == '\n') || (sent[k] == '\t') || (sent[k] == ' '));
	     --k);
	if ((k > 0) && (sent[k] == ')')) {
	    for (--k; (k >= 0) && (sent[k] != '('); --k);
	    if ((k >= 0) && (sent[k] == '(')) {
		sent[k] = '\0';

		/* Check that uttid in transcript and control file match */
		for (i = ++k;
		     sent[i] && (sent[i] != ')') &&
			 (sent[i] != '\n') && (sent[i] != '\t') && (sent[i] != ' ');
		     i++);
		sent[i] = '\0';
		if (id_cmp (sent+k, uttid) != 0)
		    E_ERROR("Uttid mismatch: ctlfile = \"%s\"; transcript = \"%s\"\n",
			   uttid, sent+k);
	    }
	}
	
	if (! feat)
	    feat = feat_array_alloc (fcb, S3_MAX_FRAMES);
	
	/* Read and process mfc/feature speech input file */
	nfr = feat_s2mfc2feat(fcb, ctlspec, cepdir, cepext, sf, ef, feat, S3_MAX_FRAMES);
	
	if (nfr <= 0){
	  E_ERROR("Utt %s: Input file read (%s) with dir (%s) and extension (%s) failed \n", uttid, ctlspec,cepdir, cepext);
	}
	else {
	    E_INFO ("%s: %d input frames\n", uttid, nfr);
	    align_utt (sent, nfr, ctlspec, uttid);
	}
	
	--ctlcount;
    }
    printf ("\n");

    while (fgets(line, sizeof(line), ctlfp) != NULL) {
	if (sscanf (line, "%s", ctlspec) > 0) {
	    E_INFO("Skipping rest of control file beginning with:\n\t%s", line);
	    break;
	}
    }

    fclose (ctlfp);
    fclose (sentfp);
    if (outsentfp)
	fclose (outsentfp);
    if (mllrctlfp)
	fclose (mllrctlfp);
}

int
main (int32 argc, char *argv[])
{
  /*  kb_t kb;
      ptmr_t tm;*/

  print_appl_info(argv[0]);
  cmd_ln_appl_enter(argc,argv,"default.arg",defn);
    
  unlimit();
  
  if ((cmd_ln_access("-mdef") == NULL) ||
      (cmd_ln_access("-mean") == NULL) ||
      (cmd_ln_access("-var") == NULL)  ||
      (cmd_ln_access("-mixw") == NULL)  ||
      (cmd_ln_access("-tmat") == NULL) ||
      (cmd_ln_access("-dict") == NULL))
    E_FATAL("Missing -mdef, -mean, -var, -mixw, -tmat, or -dict argument\n");
    
    if ((cmd_ln_access("-ctl") == NULL) || (cmd_ln_access("-insent") == NULL))
	E_FATAL("Missing -ctl or -insent argument\n");

    if ((cmd_ln_access ("-s2stsegdir") == NULL) &&
	(cmd_ln_access ("-stsegdir") == NULL) &&
	(cmd_ln_access ("-phsegdir") == NULL) &&
	(cmd_ln_access ("-wdsegdir") == NULL) &&
	(cmd_ln_access ("-outsent") == NULL))
	E_FATAL("Missing output file/directory argument(s)\n");
    
    /*
     * Initialize log(S3-base).  All scores (probs...) computed in log domain to avoid
     * underflow.  At the same time, log base = 1.0001 (1+epsilon) to allow log values
     * to be maintained in int32 variables without significant loss of precision.
     */
    if (cmd_ln_access("-logbase") == NULL)
	logs3_init (1.0001);
    else {
	float32 logbase;
	logbase = *((float32 *) cmd_ln_access("-logbase"));
	if (logbase <= 1.0)
	    E_FATAL("Illegal log-base: %e; must be > 1.0\n", logbase);
	if (logbase > 1.1)
	    E_WARN("Logbase %e perhaps too large??\n", logbase);
	logs3_init ((float64) logbase);
    }

    /*E_INFO("Log value of 3.785471 is %d\n", log_to_logs3(3.785471)); */
    /* Initialize feaure stream type */
    fcb = feat_init ( (char *) cmd_ln_access ("-feat"),
		      (char *) cmd_ln_access ("-cmn"),
		      (char *) cmd_ln_access ("-varnorm"),
		      (char *) cmd_ln_access ("-agc"));
    
    /* Read in input databases */
    models_init ();
    
    senscale = (int32 *) ckd_calloc (S3_MAX_FRAMES, sizeof(int32));
    
    timers[tmr_utt].name = "U";
    timers[tmr_gauden].name = "G";
    timers[tmr_senone].name = "S";
    timers[tmr_align].name = "A";

    /* Initialize align module */
    align_init (mdef, tmat, dict);
    printf ("\n");
    
    tot_nfr = 0;
    
    process_ctlfile ();

    if (tot_nfr > 0) {
	printf ("\n");
	printf("TOTAL FRAMES:       %8d\n", tot_nfr);
	printf("TOTAL CPU TIME:     %11.2f sec, %7.2f xRT\n",
	       tm_utt.t_tot_cpu, tm_utt.t_tot_cpu/(tot_nfr*0.01));
	printf("TOTAL ELAPSED TIME: %11.2f sec, %7.2f xRT\n",
	       tm_utt.t_tot_elapsed, tm_utt.t_tot_elapsed/(tot_nfr*0.01));
    }

#if (! WIN32)
    system ("ps aguxwww | grep s3align");
#endif

    cmd_ln_appl_exit();
    return 0;
}