cont_ad_base.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 1,062 行 · 第 1/3 页

/* ====================================================================
 * Copyright (c) 1999-2001 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.
 *
 * ====================================================================
 *
 */
/*
 * cont_ad.c -- Continuous A/D listening and silence filtering module.
 * 
 * HISTORY
 * 
 * 23-Oct-98	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Small change in the way the noiselevel is updated in find_thresh().
 * 
 * 26-Aug-98	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Separated computation of "frame power" into a separate low-level
 * 		function.
 * 
 * 13-Jul-98	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Modified to allow frame size to depend on audio sampling rate.
 * 
 * 01-Jul-98	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed CONT_AD_DELTA_SPEECH back to 20.
 * 
 * 30-Jun-98	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed CONT_AD_DELTA_SPEECH from 10 to 15.
 * 		Added FILE* argument to cont_ad_powhist_dump().
 * 
 * 19-Jun-98	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Changed CONT_AD_DELTA_SPEECH from 20 to 10, to increase sensitivity
 * 		to very short utterances.
 * 
 * 16-Jan-98	Paul Placeway ([e-mail]) at Carnegie Mellon University
 * 		Changed to use dB instead of the weird power measure.
 * 		Changed analysis window size, tuned default settings of most
 * 		parameters to make the system less sensitive to noise, changed
 * 		the histogram update frequency and decay to make the system
 * 		adapt more rapidly to changes in the environment.
 * 		Added cont_ad_set_params() and cont_ad_get_params().
 * 
 * 28-Jul-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added FRMPOW2SIGLVL, max_siglvl(), and cont_ad_t.siglvl.
 * 		Changed min signal energy/frame to CONT_AD_SPF.
 * 
 * 27-Jun-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added the option for cont_ad_read to return -1 on EOF.
 * 
 * 21-Jun-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Added cont_ad_set_thresh().
 * 		Bugfix: n_other is recomputed after updating thresholds.
 * 
 * 20-Jun-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Separated thresholds for speech and silence.
 * 		Fixed bug in moving analysis window upon transition to speech state.
 * 
 * 17-Jun-96	M K Ravishankar ([e-mail]) at Carnegie Mellon University
 * 		Created, based loosely on Steve Reed's original implementation.
 */

/*
 * This module is intended to be interposed as a filter between any raw A/D source and the
 * application to remove silence regions.  It is initialized with a raw A/D source function
 * (during the cont_ad_init call).  Filtered A/D data can be read by the application using
 * the cont_ad_read function.  This module assumes that the A/D source function supplies an
 * endless stream of data.  The application is responsible for setting up the A/D source,
 * turning recording on and off as it desires.  It is also responsible for invoking the
 * cont_ad_read function frequently enough to avoid buffer overruns and dropping A/D data.
 * This continuous listening module has an internal buffer of about 4 sec.
 * 
 * This module must be initialized and calibrated at first (cont_ad_init and cont_ad_calib
 * functions).  Raw samples are grouped into frames, the signal power in each frame is
 * computed and accumulated in a histogram.  The module is always in one of two states:
 * SILENCE or SPEECH.  Transitions between the two states are detected by looking for a
 * contiguous window of several frames that is predominantly of the other type.  The type
 * is determined by comparing frame power to either of two thresholds, thresh_sil and
 * thresh_speech, as appropriate for the current state.  These thresholds are set from the
 * first peak in the low-end of the power histogram, and are updated every few seconds.
 * Separate thresholds are used to provide some hysteresis.
 * 
 * The module maintains a linked list of speech (non-silence) segments not yet read by the
 * application.  The cont_ad_read function returns speech data, if any available, by
 * following this list.  It also updates an "absolute" timestamp at the end of the
 * cont_ad_read operation.  The timestamp indicates the total #samples of A/D data read
 * until this point, including data discarded as silence frames.  The application is
 * responsible for using this timestamp to make any policy decisions regarding utterance
 * boundaries or whatever.
 */

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

#include <s3types.h>
#include "ad.h"
#include "cont_ad.h"
#include <err.h>


#ifndef _ABS
#define _ABS(x) ((x) >= 0 ? (x) : -(x))
#endif

/* States of continuous listening module */
#define CONT_AD_STATE_SIL	0
#define CONT_AD_STATE_SPEECH	1

/* Various parameters, including defaults for many cont_ad_t member variables */

#define CONT_AD_ADFRMSIZE	256	/* #Frames of internal A/D buffer maintained */

#define CONT_AD_POWHISTSIZE	98	/* #Powhist bins: ~ FRMPOW(65536^2*CONT_AD_SPF) */
/* Maximum level is 96.3 dB full-scale; 97 for safety, plus 1 for zero-based */

#define CONT_AD_THRESH_UPDATE	100	/* Update thresholds approx every so many frames */
	/* PWP: update was 200 frames, or 3.2 seconds.  Now about every 1.6 sec. */

#define CONT_AD_ADAPT_RATE	0.2	/* Interpolation of new and old noiselevel */

#define CONT_AD_SPS             16000

#define CONT_AD_DEFAULT_NOISE	30	/* Default background noise power level */
#define CONT_AD_DELTA_SIL	5	/* Initial default for cont_ad_t.delta_sil */
#define CONT_AD_DELTA_SPEECH	20	/* Initial default for cont_ad_t.delta_speech */
#define CONT_AD_MIN_NOISE	2	/* Expected minimum background noise level */
#define CONT_AD_MAX_NOISE	70	/* Maximum background noise level */

#define CONT_AD_WINSIZE		21	/* Analysis window for state transitions */
				/* rkm had 16 */

#define CONT_AD_SPEECH_ONSET	9	/* Min #speech frames in analysis window for
					   SILENCE -> SPEECH state transition */
/*
 * SReed had 100 ms == 6.25 fr contiguous; rkm had 9 (out of 16+10) with a
 * lower threshold.
 */

#define CONT_AD_SIL_ONSET	18	/* Min #silence frames in analysis window for
					   SPEECH -> SILENCE state transition
					   MUST BE <= CONT_AD_WINSIZE */
/*
 * SReed had 400 ms == 25 fr contiguous; rkm had 14 out of 16
 */

#define CONT_AD_LEADER		5	/* On transition to SPEECH state, so many frames
					   BEFORE window included in speech data (>0) */
				/* SReed had 200 ms == 12.5 fr; rkm had 5 */

#define CONT_AD_TRAILER		10	/* On transition to SILENCE state, so many frames
					   of silence included in speech data (>0).
					   NOTE: Ensure (0 < TRAILER+LEADER <= WINSIZE) */
				/* SReed had 100 ms == 6.25 fr; rkm had 10 */

#ifdef CONT_AD_RAWDUMP
static FILE *rawfp;
#endif

static FILE *logfp = NULL;	/* Detailed info written to fp if non-NULL */
static int32 frmno = 0;

void cont_ad_powhist_dump (FILE *fp, cont_ad_t *r)
{
    int32 i;

    for (i = 0; i < CONT_AD_POWHISTSIZE; i++)
	if (r->pow_hist[i] > 0)
	    fprintf (fp, "\t%3d %6d\n", i, r->pow_hist[i]);
    fprintf (fp, "\tnew noiselevel= %d  thresh(sil,speech)= %d %d\n",
	     r->noise_level, r->thresh_sil, r->thresh_speech);
    fflush (fp);
}

/*
 * Compute frame power.  Interface deliberately kept low level to allow arbitrary
 * users to call this function with appropriate data.
 */
int32 cont_ad_frame_pow (int16 *buf, int32 *prev, int32 spf)
{
    double sumsq, v;
    int32 i;
    int32 p;
    
    sumsq = 0.0;
    p = *prev;
    for (i = 0; i < spf; i++) {
	v = (double) (buf[i] - p);
	sumsq += v*v;
	p = buf[i];
    }
    *prev = p;
    
    if (sumsq < spf)	/* Make sure FRMPOW(sumsq) >= 0 */
	sumsq = spf;

    /*
     * PWP: Units changed to dB
     *
     * Now the units of measurement of an input sample are volts (really!),
     * so the power in dB is p = 20*log10(samp).  Further, we want the RMS
     * (root-mean-squared) average power across the frame.
     *
     * "sumsq" is the sum of the sum of the squares, so we want
     *
     *   i = 20 * log10( sqrt ( sumsq / n_samps) )
     *
     * (Stephen Reed's code actually had 
     *    i = 20 * log10( sqrt (sumsq) / n_samps )
     *  but this only produced an additive error.)
     *
     * i = 20 * log10( sqrt ( sumsq / n_samps) )
     *   = 20 * log10( ( sumsq / n_samps) ^ 0.5 )
     *   = 20 * log10( ( sumsq / n_samps) ) * 0.5 )
     *   = 10 * log10( ( sumsq / n_samps) )
     *   = 10 * ( log10( sumsq) - log10(n_samps) )
     */
    i = (int32) ((10.0 * (log10(sumsq) - log10((double) spf))) + 0.5);
    if (i < 0) i = 0;		/* trim lower bound again to be safe. */
    assert (i < 97);

    return (i);
}

/*
 * Classify frame (id=frm, starting at sample position s) as sil/nonsil.  Classification
 * done in isolation, independent of any other frame, based only on power histogram.
 */
static void compute_frame_pow (cont_ad_t *r, int32 frm)
{
    int32 i;
    
    i = cont_ad_frame_pow (r->adbuf + (frm * r->spf), &(r->prev_sample), r->spf);

    if (logfp) {
        fprintf (logfp, "%8.2f %2d\n",
		 (double)(frmno * r->spf)/(double)(r->sps), i);
	fflush (logfp);
	frmno++;
    }

    r->frm_pow[frm] = (char) i;
    (r->pow_hist[i])++;
    r->thresh_update--;
}

void cont_ad_set_logfp (FILE *fp)
{
    logfp = fp;
}

/* PWP: $$$ check this */
/*
 * PWP: in SReed's code, decay was done by zeroing the histogram,
 * i.e. no history.
 */
static void decay_hist (cont_ad_t *r)
{
    int32 i;
    
    for (i = 0; i < CONT_AD_POWHISTSIZE; i++)
	r->pow_hist[i] >>= 1;
}

/*
 * Find silence threshold from power histogram.
 */
static int32 find_thresh (cont_ad_t *r)
{
    int32 i, j, max, th;

    if (!r->auto_thresh)
      return 0;

    /*
     * Find smallest non-zero histogram entry, but starting at some minimum power.
     * Power lower than CONT_AD_MIN_NOISE indicates bad A/D input (eg, mic off...).
     * Too high a minimum power is also bad.
     */
    for (i = r->min_noise; (i < CONT_AD_POWHISTSIZE) && (r->pow_hist[i] == 0); i++);
    if (i > r->max_noise)	/* Bad signal? */
	return -1;

    /* PWP: Hmmmmm.... SReed's code looks over the lower 20 dB */
    /* PWP: 1/14/98  Made to work like Stephen Reed's code */

    max = 0;
    for (j = i, th = i;
	 (j < CONT_AD_POWHISTSIZE) && (j < i+20); j++) { /* PWP: was i+6, which was 9 dB */
	if (max < r->pow_hist[j]) {
	    max = r->pow_hist[j];
	    th = j;
	}
    }

    if (logfp) {
	fprintf (logfp, "\tfrm= %d  noiselevel= %d  histthresh= %d\n",
		 frmno, r->noise_level, th);
    }
    
    /* "Don't change the threshold too fast" */
#if 0
    if ( _ABS(r->noise_level - th) >= 10 ) {
	if (th > r->noise_level)
	    r->noise_level += ((th - r->noise_level) / 2);
	else
	    r->noise_level -= ((r->noise_level - th) / 2);
    } else {
	r->noise_level = th;
    }
#else
    /*
     * RKM: The above is odd; if (diff >= 10) += diff/2; else += diff??
     * This is discontinuous.  Change to always += diff/2.
     */
    r->noise_level = (int32) (th * r->adapt_rate + r->noise_level * (1.0 - r->adapt_rate));
#endif

    /* update thresholds */
    r->thresh_sil = r->noise_level + r->delta_sil;
    r->thresh_speech = r->noise_level + r->delta_speech;

    if (logfp)
	cont_ad_powhist_dump (logfp, r);
    
    /*