flat_fwd.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 2,212 行 · 第 1/5 页

	    printf ("[%4d] %-24s", n_frm, dict->word[w].word);

	    last = dict->word[w].pronlen - 1;
	    bestlast = (int32) 0x80000000;
	    
	    for (h = whmm[w]; h; h = h->next) {
		if (h->pos < last)
		    printf (" %9d.%2d", -h->score[n_state-1], h->pos);
		else if (bestlast < h->score[n_state-1])
		    bestlast = h->score[n_state-1];
	    }

	    if (bestlast > (int32) 0x80000000)
		printf (" %9d.%2d", -bestlast, last);

	    printf ("\n");
	}
    }
}


/**
 * Check model tprob matrices that they conform to upper-diagonal assumption.
 */
static void chk_tp_uppertri ( void )
{
    int32 i, from, to;
    
    assert (n_state > 0);
    
    /* Check that each tmat is upper-triangular */
    for (i = 0; i < tmat->n_tmat; i++) {
	for (to = 0; to < n_state-1; to++)
	    for (from = to+1; from < n_state-1; from++)
		if (tmat->tp[i][from][to] > S3_LOGPROB_ZERO)
		    E_FATAL("HMM transition matrix not upper triangular\n");
    }
}


/** For partial evaluation of incoming state score (prev state score + senone score) */
static int32 *st_sen_scr;

#define ANYHMMTOPO 1
#if (! ANYHMMTOPO)
/**
 * Like the general eval_nonmpx_whmm and eval_mpx_whmm below, but hardwired for
 * the Sphinx-II 5-state Bakis topology.
 */
static void eval_nonmpx_whmm (s3wid_t w, whmm_t *h, int32 *senscr)
{
    register int32 s0, s1, s2, s3, s4;
    register int32 scr, newscr1, newscr2, bestscr;
    register int32 *tp;
    s3pid_t p;
    s3senid_t *senp;

    p = *(h->pid);
    senp = mdef->phone[p].state;
    tp = tmat->tp[mdef->phone[p].tmat][0];	/* HACK!! Assumes tp 2-D data allocated
						   contiguously */
    
    if ((s0 = h->score[0] + senscr[senp[0]]) < S3_LOGPROB_ZERO)
	s0 = S3_LOGPROB_ZERO;
    if ((s1 = h->score[1] + senscr[senp[1]]) < S3_LOGPROB_ZERO)
	s1 = S3_LOGPROB_ZERO;
    if ((s2 = h->score[2] + senscr[senp[2]]) < S3_LOGPROB_ZERO)
	s2 = S3_LOGPROB_ZERO;
    if ((s3 = h->score[3] + senscr[senp[3]]) < S3_LOGPROB_ZERO)
	s3 = S3_LOGPROB_ZERO;
    if ((s4 = h->score[4] + senscr[senp[4]]) < S3_LOGPROB_ZERO)
	s4 = S3_LOGPROB_ZERO;
    
    newscr1 = s4 + tp[29]; /* [4][5] */
    newscr2 = s3 + tp[23]; /* [3][5] */
    if (newscr1 > newscr2) {
	h->score[5] = bestscr = newscr1;
	h->history[5] = h->history[4];
    } else {
	h->score[5] = bestscr = newscr2;
	h->history[5] = h->history[3];
    }

    scr     = s4 + tp[28]; /* [4][4] */
    newscr1 = s3 + tp[22]; /* [3][4] */
    newscr2 = s2 + tp[16]; /* [2][4] */
    if (newscr2 > newscr1) {
	if (newscr2 > scr) {
	    h->score[4] = newscr2;
	    h->history[4] = h->history[2];
	} else
	    h->score[4] = scr;
    } else {
	if (newscr1 > scr) {
	    h->score[4] = newscr1;
	    h->history[4] = h->history[3];
	} else
	    h->score[4] = scr;
    }
    if (bestscr < h->score[4])
	bestscr = h->score[4];
    
    scr     = s3 + tp[21]; /* [3][3] */
    newscr1 = s2 + tp[15]; /* [2][3] */
    newscr2 = s1 + tp[9];  /* [1][3] */
    if (newscr2 > newscr1) {
	if (newscr2 > scr) {
	    h->score[3] = newscr2;
	    h->history[3] = h->history[1];
	} else
	    h->score[3] = scr;
    } else {
	if (newscr1 > scr) {
	    h->score[3] = newscr1;
	    h->history[3] = h->history[2];
	} else
	    h->score[3] = scr;
    }
    if (bestscr < h->score[3])
	bestscr = h->score[3];
    
    scr     = s2 + tp[14]; /* [2][2] */
    newscr1 = s1 + tp[8];  /* [1][2] */
    newscr2 = s0 + tp[2];  /* [0][2] */
    if (newscr2 > newscr1) {
	if (newscr2 > scr) {
	    h->score[2] = newscr2;
	    h->history[2] = h->history[0];
	} else
	    h->score[2] = scr;
    } else {
	if (newscr1 > scr) {
	    h->score[2] = newscr1;
	    h->history[2] = h->history[1];
	} else
	    h->score[2] = scr;
    }
    if (bestscr < h->score[2])
	bestscr = h->score[2];
    
    scr     = s1 + tp[7]; /* [1][1] */
    newscr1 = s0 + tp[1]; /* [0][1] */
    if (newscr1 > scr) {
	h->score[1] = newscr1;
	h->history[1] = h->history[0];
    } else
	h->score[1] = scr;
    if (bestscr < h->score[1])
	bestscr = h->score[1];
    
    h->score[0] = scr = s0 + tp[0];	/* [0][0] */
    if (bestscr < scr)
	bestscr = scr;
    
    h->bestscore = bestscr;
}


static void eval_mpx_whmm (s3wid_t w, whmm_t *h, int32 *senscr)
{
    register int32 s0, s1, s2, s3, s4;
    register int32 *tp0, *tp1, *tp2, *tp3, *tp4;
    register int32 scr, newscr1, newscr2, bestscr;
    s3senid_t *senp;
    s3pid_t p0, p1, p2, p3, p4;

    p0 = h->pid[0];
    p1 = h->pid[1];
    p2 = h->pid[2];
    p3 = h->pid[3];
    p4 = h->pid[4];

    senp = mdef->phone[p0].state;
    if ((s0 = h->score[0] + senscr[senp[0]]) < S3_LOGPROB_ZERO)
	s0 = S3_LOGPROB_ZERO;
    tp0 = tmat->tp[mdef->phone[p0].tmat][0];	/* HACK!! See eval_nonmpx_whmm */

    if (p1 != p0) {
	senp = mdef->phone[p1].state;
	tp1 = tmat->tp[mdef->phone[p1].tmat][0];/* HACK!! See eval_nonmpx_whmm */
    } else
	tp1 = tp0;
    if ((s1 = h->score[1] + senscr[senp[1]]) < S3_LOGPROB_ZERO)
	s1 = S3_LOGPROB_ZERO;

    if (p2 != p1) {
	senp = mdef->phone[p2].state;
	tp2 = tmat->tp[mdef->phone[p2].tmat][0];/* HACK!! See eval_nonmpx_whmm */
    } else
	tp2 = tp1;
    if ((s2 = h->score[2] + senscr[senp[2]]) < S3_LOGPROB_ZERO)
	s2 = S3_LOGPROB_ZERO;

    if (p3 != p2) {
	senp = mdef->phone[p3].state;
	tp3 = tmat->tp[mdef->phone[p3].tmat][0];/* HACK!! See eval_nonmpx_whmm */
    } else
	tp3 = tp2;
    if ((s3 = h->score[3] + senscr[senp[3]]) < S3_LOGPROB_ZERO)
	s3 = S3_LOGPROB_ZERO;

    if (p4 != p3) {
	senp = mdef->phone[p4].state;
	tp4 = tmat->tp[mdef->phone[p4].tmat][0];/* HACK!! See eval_nonmpx_whmm */
    } else
	tp4 = tp3;
    if ((s4 = h->score[4] + senscr[senp[4]]) < S3_LOGPROB_ZERO)
	s4 = S3_LOGPROB_ZERO;
    
    newscr1 = s4 + tp4[29]; /* [4][5] */
    newscr2 = s3 + tp3[23]; /* [3][5] */
    if (newscr1 > newscr2) {
	h->score[5] = bestscr = newscr1;
	h->history[5] = h->history[4];
    } else {
	h->score[5] = bestscr = newscr2;
	h->history[5] = h->history[3];
    }

    scr     = s4 + tp4[28]; /* [4][4] */
    newscr1 = s3 + tp3[22]; /* [3][4] */
    newscr2 = s2 + tp2[16]; /* [2][4] */
    if (newscr2 > newscr1) {
	if (newscr2 > scr) {
	    h->score[4] = newscr2;
	    h->history[4] = h->history[2];
	    h->pid[4] = h->pid[2];
	} else
	    h->score[4] = scr;
    } else {
	if (newscr1 > scr) {
	    h->score[4] = newscr1;
	    h->history[4] = h->history[3];
	    h->pid[4] = h->pid[3];
	} else
	    h->score[4] = scr;
    }
    if (bestscr < h->score[4])
	bestscr = h->score[4];
    
    scr     = s3 + tp3[21]; /* [3][3] */
    newscr1 = s2 + tp2[15]; /* [2][3] */
    newscr2 = s1 + tp1[9];  /* [1][3] */
    if (newscr2 > newscr1) {
	if (newscr2 > scr) {
	    h->score[3] = newscr2;
	    h->history[3] = h->history[1];
	    h->pid[3] = h->pid[1];
	} else
	    h->score[3] = scr;
    } else {
	if (newscr1 > scr) {
	    h->score[3] = newscr1;
	    h->history[3] = h->history[2];
	    h->pid[3] = h->pid[2];
	} else
	    h->score[3] = scr;
    }
    if (bestscr < h->score[3])
	bestscr = h->score[3];
    
    scr     = s2 + tp2[14]; /* [2][2] */
    newscr1 = s1 + tp1[8];  /* [1][2] */
    newscr2 = s0 + tp0[2];  /* [0][2] */
    if (newscr2 > newscr1) {
	if (newscr2 > scr) {
	    h->score[2] = newscr2;
	    h->history[2] = h->history[0];
	    h->pid[2] = h->pid[0];
	} else
	    h->score[2] = scr;
    } else {
	if (newscr1 > scr) {
	    h->score[2] = newscr1;
	    h->history[2] = h->history[1];
	    h->pid[2] = h->pid[1];
	} else
	    h->score[2] = scr;
    }
    if (bestscr < h->score[2])
	bestscr = h->score[2];
    
    scr     = s1 + tp1[7]; /* [1][1] */
    newscr1 = s0 + tp0[1]; /* [0][1] */
    if (newscr1 > scr) {
	h->score[1] = newscr1;
	h->history[1] = h->history[0];
	h->pid[1] = h->pid[0];
    } else
	h->score[1] = scr;
    if (bestscr < h->score[1])
	bestscr = h->score[1];
    
    h->score[0] = scr = s0 + tp0[0]; /* [0][0] */
    if (bestscr < scr)
	bestscr = scr;
    
    h->bestscore = bestscr;
}

#else

/**
 * Evaluate non-multiplexed word HMM (ie, the entire whmm really represents one
 * phone rather than each state representing a potentially different phone.
 */
static void eval_nonmpx_whmm (s3wid_t w, whmm_t *h, int32 *senscr)
{
    s3pid_t pid;
    s3senid_t *sen;
    int32 **tp;
    int32 to, from, bestfrom;
    int32 newscr, scr, bestscr;

    pid = *(h->pid);
    
    sen = mdef->phone[pid].state;
    tp = tmat->tp[mdef->phone[pid].tmat];

    /* Compute previous state-score + observation output prob for each state */
    for (from = n_state-2; from >= 0; --from) {
	if ((st_sen_scr[from] = h->score[from] + senscr[sen[from]]) < S3_LOGPROB_ZERO)
	    st_sen_scr[from] = S3_LOGPROB_ZERO;
    }
    
    /* Evaluate final-state first, which does not have a self-transition */
    to = final_state;
    scr = S3_LOGPROB_ZERO;
    bestfrom = -1;
    for (from = to-1; from >= 0; --from) {
	if ((tp[from][to] > S3_LOGPROB_ZERO) &&
	    ((newscr = st_sen_scr[from] + tp[from][to]) > scr)) {
	    scr = newscr;
	    bestfrom = from;
	}
    }
    h->score[to] = scr;
    if (bestfrom >= 0)
	h->history[to] = h->history[bestfrom];

    bestscr = scr;

    /* Evaluate all other states, which might have self-transitions */
    for (to = final_state-1; to >= 0; --to) {
	/* Score from self-transition, if any */
	scr = (tp[to][to] > S3_LOGPROB_ZERO) ? st_sen_scr[to] + tp[to][to] : S3_LOGPROB_ZERO;

	/* Scores from transitions from other states */
	bestfrom = -1;
	for (from = to-1; from >= 0; --from) {
	    if ((tp[from][to] > S3_LOGPROB_ZERO) &&
		((newscr = st_sen_scr[from] + tp[from][to]) > scr)) {
		scr = newscr;
		bestfrom = from;
	    }
	}

	/* Update new result for state to */
	h->score[to] = scr;
	if (bestfrom >= 0)
	    h->history[to] = h->history[bestfrom];

	if (bestscr < scr)
	    bestscr = scr;
    }

    h->bestscore = bestscr;
}


/** Like eval_nonmpx_whmm, except there's a different pid associated with each state */
static void eval_mpx_whmm (s3wid_t w, whmm_t *h, int32 *senscr)
{
    s3pid_t pid, prevpid;
    s3senid_t *senp;
    int32 **tp;
    int32 to, from, bestfrom;
    int32 newscr, scr, bestscr;
    
    senp=NULL;
    tp=NULL;
    /* Compute previous state-score + observation output prob for each state */
    prevpid = BAD_S3PID;
    for (from = n_state-2; from >= 0; --from) {
	if ((pid = h->pid[from]) != prevpid) {
	    senp = mdef->phone[pid].state;
	    prevpid = pid;
	}

	if ((st_sen_scr[from] = h->score[from] + senscr[senp[from]]) < S3_LOGPROB_ZERO)
	    st_sen_scr[from] = S3_LOGPROB_ZERO;
    }

    /* Evaluate final-state first, which does not have a self-transition */
    to = final_state;
    scr = S3_LOGPROB_ZERO;
    bestfrom = -1;
    prevpid = BAD_S3PID;
    for (from = to-1; from >= 0; --from) {
	if ((pid = h->pid[from]) != prevpid) {
	    tp = tmat->tp[mdef->phone[pid].tmat];
	    prevpid = pid;
	}

	if ((tp[from][to] > S3_LOGPROB_ZERO) &&
	    ((newscr = st_sen_scr[from] + tp[from][to]) > scr)) {
	    scr = newscr;
	    bestfrom = from;
	}
    }
    h->score[to] = scr;
    if (bestfrom >= 0) {
	h->history[to] = h->history[bestfrom];
	h->pid[to] = h->pid[bestfrom];
    }
    
    bestscr = scr;

    /* Evaluate all other states, which might have self-transitions */
    for (to = final_state-1; to >= 0; --to) {
	/* Score from self-transition, if any */
	if ((pid = h->pid[to]) != prevpid) {
	    tp = tmat->tp[mdef->phone[pid].tmat];
	    prevpid = pid;
	}
	scr = (tp[to][to] > S3_LOGPROB_ZERO) ? st_sen_scr[to] + tp[to][to] : S3_LOGPROB_ZERO;

	/* Scores from transitions from other states */
	bestfrom = -1;
	for (from = to-1; from >= 0; --from) {
	    if ((pid = h->pid[from]) != prevpid) {
		tp = tmat->tp[mdef->phone[pid].tmat];
		prevpid = pid;
	    }
	    
	    if ((tp[from][to] > S3_LOGPROB_ZERO) &&
		((newscr = st_sen_scr[from] + tp[from][to]) > scr)) {
		scr = newscr;
		bestfrom = from;
	    }
	}

	/* Update new result for state to */