tenfourfox/media/pocketsphinx/src/ngram_search.c

/* -*- c-basic-offset: 4; indent-tabs-mode: nil -*- */
/* ====================================================================
 * Copyright (c) 2008 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.
 *
 * ====================================================================
 *
 */

/**
 * @file ngram_search.c N-Gram based multi-pass search ("FBS")
 */

/* System headers. */
#include <string.h>
#include <assert.h>

/* SphinxBase headers. */
#include <sphinxbase/ckd_alloc.h>
#include <sphinxbase/listelem_alloc.h>
#include <sphinxbase/err.h>

/* Local headers. */
#include "pocketsphinx_internal.h"
#include "ps_lattice_internal.h"
#include "ngram_search.h"
#include "ngram_search_fwdtree.h"
#include "ngram_search_fwdflat.h"

static int ngram_search_start(ps_search_t *search);
static int ngram_search_step(ps_search_t *search, int frame_idx);
static int ngram_search_finish(ps_search_t *search);
static int ngram_search_reinit(ps_search_t *search, dict_t *dict, dict2pid_t *d2p);
static char const *ngram_search_hyp(ps_search_t *search, int32 *out_score, int32 *out_is_final);
static int32 ngram_search_prob(ps_search_t *search);
static ps_seg_t *ngram_search_seg_iter(ps_search_t *search, int32 *out_score);

static ps_searchfuncs_t ngram_funcs = {
    /* name: */   "ngram",
    /* start: */  ngram_search_start,
    /* step: */   ngram_search_step,
    /* finish: */ ngram_search_finish,
    /* reinit: */ ngram_search_reinit,
    /* free: */   ngram_search_free,
    /* lattice: */  ngram_search_lattice,
    /* hyp: */      ngram_search_hyp,
    /* prob: */     ngram_search_prob,
    /* seg_iter: */ ngram_search_seg_iter,
};

static ngram_model_t *default_lm;

static void
ngram_search_update_widmap(ngram_search_t *ngs)
{
    char const **words;
    int32 i, n_words;

    /* It's okay to include fillers since they won't be in the LM */
    n_words = ps_search_n_words(ngs);
    words = (char const**)ckd_calloc(n_words, sizeof(*words));
    /* This will include alternates, again, that's okay since they aren't in the LM */
    for (i = 0; i < n_words; ++i)
        words[i] = dict_wordstr(ps_search_dict(ngs), i);
    ngram_model_set_map_words(ngs->lmset, words, n_words);
    ckd_free(words);
}

static void
ngram_search_calc_beams(ngram_search_t *ngs)
{
    cmd_ln_t *config;
    acmod_t *acmod;

    config = ps_search_config(ngs);
    acmod = ps_search_acmod(ngs);

    /* Log beam widths. */
    ngs->beam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-beam"))>>SENSCR_SHIFT;
    ngs->wbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-wbeam"))>>SENSCR_SHIFT;
    ngs->pbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pbeam"))>>SENSCR_SHIFT;
    ngs->lpbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-lpbeam"))>>SENSCR_SHIFT;
    ngs->lponlybeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-lponlybeam"))>>SENSCR_SHIFT;
    ngs->fwdflatbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-fwdflatbeam"))>>SENSCR_SHIFT;
    ngs->fwdflatwbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-fwdflatwbeam"))>>SENSCR_SHIFT;

    /* Absolute pruning parameters. */
    ngs->maxwpf = cmd_ln_int32_r(config, "-maxwpf");
    ngs->maxhmmpf = cmd_ln_int32_r(config, "-maxhmmpf");

    /* Various penalties which may or may not be useful. */
    ngs->wip = logmath_log(acmod->lmath, cmd_ln_float32_r(config, "-wip")) >>SENSCR_SHIFT;
    ngs->nwpen = logmath_log(acmod->lmath, cmd_ln_float32_r(config, "-nwpen")) >>SENSCR_SHIFT;
    ngs->pip = logmath_log(acmod->lmath, cmd_ln_float32_r(config, "-pip")) >>SENSCR_SHIFT;
    ngs->silpen = ngs->pip
        + (logmath_log(acmod->lmath, cmd_ln_float32_r(config, "-silprob"))>>SENSCR_SHIFT);
    ngs->fillpen = ngs->pip
        + (logmath_log(acmod->lmath, cmd_ln_float32_r(config, "-fillprob"))>>SENSCR_SHIFT);

    /* Language weight ratios for fwdflat and bestpath search. */
    ngs->fwdflat_fwdtree_lw_ratio =
        cmd_ln_float32_r(config, "-fwdflatlw")
        / cmd_ln_float32_r(config, "-lw");
    ngs->bestpath_fwdtree_lw_ratio =
        cmd_ln_float32_r(config, "-bestpathlw")
        / cmd_ln_float32_r(config, "-lw");

    /* Acoustic score scale for posterior probabilities. */
    ngs->ascale = 1.0 / cmd_ln_float32_r(config, "-ascale");
}

ps_search_t *
ngram_search_init(ngram_model_t *lm,
                  cmd_ln_t *config,
                  acmod_t *acmod,
                  dict_t *dict,
                  dict2pid_t *d2p)
{
    ngram_search_t *ngs;
    static char *lmname = "default";

    /* Make the acmod's feature buffer growable if we are doing two-pass
     * search. */
    acmod_set_grow(acmod, cmd_ln_boolean_r(config, "-fwdflat") &&
                          cmd_ln_boolean_r(config, "-fwdtree"));

    ngs = ckd_calloc(1, sizeof(*ngs));
    ps_search_init(&ngs->base, &ngram_funcs, config, acmod, dict, d2p);
    ngs->hmmctx = hmm_context_init(bin_mdef_n_emit_state(acmod->mdef),
                                   acmod->tmat->tp, NULL, acmod->mdef->sseq);
    if (ngs->hmmctx == NULL) {
        ps_search_free(ps_search_base(ngs));
        return NULL;
    }
    ngs->chan_alloc = listelem_alloc_init(sizeof(chan_t));
    ngs->root_chan_alloc = listelem_alloc_init(sizeof(root_chan_t));
    ngs->latnode_alloc = listelem_alloc_init(sizeof(ps_latnode_t));

    /* Calculate various beam widths and such. */
    ngram_search_calc_beams(ngs);

    /* Allocate a billion different tables for stuff. */
    ngs->word_chan = ckd_calloc(dict_size(dict),
                                sizeof(*ngs->word_chan));
    ngs->word_lat_idx = ckd_calloc(dict_size(dict),
                                   sizeof(*ngs->word_lat_idx));
    ngs->word_active = bitvec_alloc(dict_size(dict));
    ngs->last_ltrans = ckd_calloc(dict_size(dict),
                                  sizeof(*ngs->last_ltrans));

    /* FIXME: All these structures need to be made dynamic with
     * garbage collection. */
    ngs->bp_table_size = cmd_ln_int32_r(config, "-latsize");
    ngs->bp_table = ckd_calloc(ngs->bp_table_size,
                               sizeof(*ngs->bp_table));
    /* FIXME: This thing is frickin' huge. */
    ngs->bscore_stack_size = ngs->bp_table_size * 20;
    ngs->bscore_stack = ckd_calloc(ngs->bscore_stack_size,
                                   sizeof(*ngs->bscore_stack));
    ngs->n_frame_alloc = 256;
    ngs->bp_table_idx = ckd_calloc(ngs->n_frame_alloc + 1,
                                   sizeof(*ngs->bp_table_idx));
    ++ngs->bp_table_idx; /* Make bptableidx[-1] valid */

    /* Allocate active word list array */
    ngs->active_word_list = ckd_calloc_2d(2, dict_size(dict),
                                          sizeof(**ngs->active_word_list));

    ngs->lmset = ngram_model_set_init(config, &lm, &lmname, NULL, 1);
    if (!ngs->lmset)
        goto error_out;

    if (ngram_wid(ngs->lmset, S3_FINISH_WORD) ==
        ngram_unknown_wid(ngs->lmset))
    {
        E_ERROR("Language model/set does not contain </s>, "
                "recognition will fail\n");
        goto error_out;
    }

    /* Create word mappings. */
    ngram_search_update_widmap(ngs);

    /* Initialize fwdtree, fwdflat, bestpath modules if necessary. */
    if (cmd_ln_boolean_r(config, "-fwdtree")) {
        ngram_fwdtree_init(ngs);
        ngs->fwdtree = TRUE;
        ngs->fwdtree_perf.name = "fwdtree";
        ptmr_init(&ngs->fwdtree_perf);
    }
    if (cmd_ln_boolean_r(config, "-fwdflat")) {
        ngram_fwdflat_init(ngs);
        ngs->fwdflat = TRUE;
        ngs->fwdflat_perf.name = "fwdflat";
        ptmr_init(&ngs->fwdflat_perf);
    }
    if (cmd_ln_boolean_r(config, "-bestpath")) {
        ngs->bestpath = TRUE;
        ngs->bestpath_perf.name = "bestpath";
        ptmr_init(&ngs->bestpath_perf);
    }

    return (ps_search_t *)ngs;

error_out:
    ngram_search_free((ps_search_t *)ngs);
    return NULL;
}

static int
ngram_search_reinit(ps_search_t *search, dict_t *dict, dict2pid_t *d2p)
{
    ngram_search_t *ngs = (ngram_search_t *)search;
    int old_n_words;
    int rv = 0;

    /* Update the number of words. */
    old_n_words = search->n_words;
    if (old_n_words != dict_size(dict)) {
        search->n_words = dict_size(dict);
        /* Reallocate these temporary arrays. */
        ckd_free(ngs->word_lat_idx);
        ckd_free(ngs->word_active);
        ckd_free(ngs->last_ltrans);
        ckd_free_2d(ngs->active_word_list);
        ngs->word_lat_idx = ckd_calloc(search->n_words, sizeof(*ngs->word_lat_idx));
        ngs->word_active = bitvec_alloc(search->n_words);
        ngs->last_ltrans = ckd_calloc(search->n_words, sizeof(*ngs->last_ltrans));
        ngs->active_word_list
            = ckd_calloc_2d(2, search->n_words,
                            sizeof(**ngs->active_word_list));
    }

    /* Free old dict2pid, dict */
    ps_search_base_reinit(search, dict, d2p);

    if (ngs->lmset == NULL)
        return 0;

    /* Update beam widths. */
    ngram_search_calc_beams(ngs);

    /* Update word mappings. */
    ngram_search_update_widmap(ngs);

    /* Now rebuild lextrees. */
    if (ngs->fwdtree) {
        if ((rv = ngram_fwdtree_reinit(ngs)) < 0)
            return rv;
    }
    if (ngs->fwdflat) {
        if ((rv = ngram_fwdflat_reinit(ngs)) < 0)
            return rv;
    }

    return rv;
}

void
ngram_search_free(ps_search_t *search)
{
    ngram_search_t *ngs = (ngram_search_t *)search;

    ps_search_deinit(search);
    if (ngs->fwdtree)
        ngram_fwdtree_deinit(ngs);
    if (ngs->fwdflat)
        ngram_fwdflat_deinit(ngs);
    if (ngs->bestpath) {
        double n_speech = (double)ngs->n_tot_frame
            / cmd_ln_int32_r(ps_search_config(ngs), "-frate");

        E_INFO("TOTAL bestpath %.2f CPU %.3f xRT\n",
               ngs->bestpath_perf.t_tot_cpu,
               ngs->bestpath_perf.t_tot_cpu / n_speech);
        E_INFO("TOTAL bestpath %.2f wall %.3f xRT\n",
               ngs->bestpath_perf.t_tot_elapsed,
               ngs->bestpath_perf.t_tot_elapsed / n_speech);
    }

    hmm_context_free(ngs->hmmctx);
    listelem_alloc_free(ngs->chan_alloc);
    listelem_alloc_free(ngs->root_chan_alloc);
    listelem_alloc_free(ngs->latnode_alloc);
    ngram_model_free(ngs->lmset);

    ckd_free(ngs->word_chan);
    ckd_free(ngs->word_lat_idx);
    bitvec_free(ngs->word_active);
    ckd_free(ngs->bp_table);
    ckd_free(ngs->bscore_stack);
    if (ngs->bp_table_idx != NULL)
        ckd_free(ngs->bp_table_idx - 1);
    ckd_free_2d(ngs->active_word_list);
    ckd_free(ngs->last_ltrans);
    ckd_free(ngs);
}

int
ngram_search_mark_bptable(ngram_search_t *ngs, int frame_idx)
{
    if (frame_idx >= ngs->n_frame_alloc) {
        ngs->n_frame_alloc *= 2;
        ngs->bp_table_idx = ckd_realloc(ngs->bp_table_idx - 1,
                                        (ngs->n_frame_alloc + 1)
                                        * sizeof(*ngs->bp_table_idx));
        if (ngs->frm_wordlist) {
            ngs->frm_wordlist = ckd_realloc(ngs->frm_wordlist,
                                            ngs->n_frame_alloc
                                            * sizeof(*ngs->frm_wordlist));
        }
        ++ngs->bp_table_idx; /* Make bptableidx[-1] valid */
    }
    ngs->bp_table_idx[frame_idx] = ngs->bpidx;
    return ngs->bpidx;
}

static void
set_real_wid(ngram_search_t *ngs, int32 bp)
{
    bptbl_t *ent, *prev;

    assert(bp != NO_BP);
    ent = ngs->bp_table + bp;
    if (ent->bp == NO_BP)
        prev = NULL;
    else
        prev = ngs->bp_table + ent->bp;

    /* Propagate lm state for fillers, rotate it for words. */
    if (dict_filler_word(ps_search_dict(ngs), ent->wid)) {
        if (prev != NULL) {
            ent->real_wid = prev->real_wid;
            ent->prev_real_wid = prev->prev_real_wid;
        }
        else {
            ent->real_wid = dict_basewid(ps_search_dict(ngs),
                                         ent->wid);
            ent->prev_real_wid = BAD_S3WID;
        }
    }
    else {
        ent->real_wid = dict_basewid(ps_search_dict(ngs), ent->wid);
        if (prev != NULL)
            ent->prev_real_wid = prev->real_wid;
        else
            ent->prev_real_wid = BAD_S3WID;
    }
}

#define NGRAM_HISTORY_LONG_WORD 2000 /* 20s */

void
ngram_search_save_bp(ngram_search_t *ngs, int frame_idx,
                     int32 w, int32 score, int32 path, int32 rc)
{
    int32 bp;

    /* Look for an existing exit for this word in this frame.  The
     * only reason one would exist is from a different right context
     * triphone, but of course that happens quite frequently. */
    bp = ngs->word_lat_idx[w];
    if (bp != NO_BP) {

        if (frame_idx - ngs->bp_table[path].frame > NGRAM_HISTORY_LONG_WORD) {
    	    E_WARN("Word '%s' survived for %d frames, potential overpruning\n", dict_wordstr(ps_search_dict(ngs), w),
	    	    frame_idx - ngs->bp_table[path].frame);
	}

        /* Keep only the best scoring one, we will reconstruct the
         * others from the right context scores - usually the history
         * is not lost. */
        if (ngs->bp_table[bp].score WORSE_THAN score) {
            assert(path != bp); /* Pathological. */
            if (ngs->bp_table[bp].bp != path) {
                int32 bplh[2], newlh[2];
                /* But, sometimes, the history *is* lost.  If we wanted to
                 * do exact language model scoring we'd have to preserve
                 * these alternate histories. */
                E_DEBUG(2,("Updating path history %d => %d frame %d\n",
                           ngs->bp_table[bp].bp, path, frame_idx));
                bplh[0] = ngs->bp_table[bp].bp == -1
                    ? -1 : ngs->bp_table[ngs->bp_table[bp].bp].prev_real_wid;
                bplh[1] = ngs->bp_table[bp].bp == -1
                    ? -1 : ngs->bp_table[ngs->bp_table[bp].bp].real_wid;
                newlh[0] = path == -1
                    ? -1 : ngs->bp_table[path].prev_real_wid;
                newlh[1] = path == -1
                    ? -1 : ngs->bp_table[path].real_wid;
                /* Actually it's worth checking how often the actual
                 * language model state changes. */
                if (bplh[0] != newlh[0] || bplh[1] != newlh[1]) {
                    /* It's fairly rare that the actual language model
                     * state changes, but it does happen some
                     * times. */
                    E_DEBUG(1, ("Updating language model state %s,%s => %s,%s frame %d\n",
                                dict_wordstr(ps_search_dict(ngs), bplh[0]),
                                dict_wordstr(ps_search_dict(ngs), bplh[1]),
                                dict_wordstr(ps_search_dict(ngs), newlh[0]),
                                dict_wordstr(ps_search_dict(ngs), newlh[1]),
                                frame_idx));
                    set_real_wid(ngs, bp);
                }
                ngs->bp_table[bp].bp = path;
            }
            ngs->bp_table[bp].score = score;
        }
        /* But do keep track of scores for all right contexts, since
         * we need them to determine the starting path scores for any
         * successors of this word exit. */
        if (ngs->bp_table[bp].s_idx != -1)
            ngs->bscore_stack[ngs->bp_table[bp].s_idx + rc] = score;
    }
    else {
        int32 i, rcsize;
        bptbl_t *be;

        /* This might happen if recognition fails. */
        if (ngs->bpidx == NO_BP) {
            E_ERROR("No entries in backpointer table!");
            return;
        }

        /* Expand the backpointer tables if necessary. */
        if (ngs->bpidx >= ngs->bp_table_size) {
            ngs->bp_table_size *= 2;
            ngs->bp_table = ckd_realloc(ngs->bp_table,
                                        ngs->bp_table_size
                                        * sizeof(*ngs->bp_table));
            E_INFO("Resized backpointer table to %d entries\n", ngs->bp_table_size);
        }
        if (ngs->bss_head >= ngs->bscore_stack_size
            - bin_mdef_n_ciphone(ps_search_acmod(ngs)->mdef)) {
            ngs->bscore_stack_size *= 2;
            ngs->bscore_stack = ckd_realloc(ngs->bscore_stack,
                                            ngs->bscore_stack_size
                                            * sizeof(*ngs->bscore_stack));
            E_INFO("Resized score stack to %d entries\n", ngs->bscore_stack_size);
        }

        ngs->word_lat_idx[w] = ngs->bpidx;
        be = &(ngs->bp_table[ngs->bpidx]);
        be->wid = w;
        be->frame = frame_idx;
        be->bp = path;
        be->score = score;
        be->s_idx = ngs->bss_head;
        be->valid = TRUE;
        assert(path != ngs->bpidx);

        /* DICT2PID */
        /* Get diphone ID for final phone and number of ssids corresponding to it. */
        be->last_phone = dict_last_phone(ps_search_dict(ngs),w);
        if (dict_is_single_phone(ps_search_dict(ngs), w)) {
            be->last2_phone = -1;
            be->s_idx = -1;
            rcsize = 0;
        }
        else {
            be->last2_phone = dict_second_last_phone(ps_search_dict(ngs),w);
            rcsize = dict2pid_rssid(ps_search_dict2pid(ngs),
                                    be->last_phone, be->last2_phone)->n_ssid;
        }
        /* Allocate some space on the bscore_stack for all of these triphones. */
        for (i = 0; i < rcsize; ++i)
            ngs->bscore_stack[ngs->bss_head + i] = WORST_SCORE;
        if (rcsize)
            ngs->bscore_stack[ngs->bss_head + rc] = score;
        set_real_wid(ngs, ngs->bpidx);

        ngs->bpidx++;
        ngs->bss_head += rcsize;
    }
}

int
ngram_search_find_exit(ngram_search_t *ngs, int frame_idx, int32 *out_best_score, int32 *out_is_final)
{
    /* End of backpointers for this frame. */
    int end_bpidx;
    int best_exit, bp;
    int32 best_score;

    /* No hypothesis means no exit node! */
    if (ngs->n_frame == 0)
        return NO_BP;

    if (frame_idx == -1 || frame_idx >= ngs->n_frame)
        frame_idx = ngs->n_frame - 1;
    end_bpidx = ngs->bp_table_idx[frame_idx];

    best_score = WORST_SCORE;
    best_exit = NO_BP;

    /* Scan back to find a frame with some backpointers in it. */
    while (frame_idx >= 0 && ngs->bp_table_idx[frame_idx] == end_bpidx)
        --frame_idx;
    /* This is NOT an error, it just means there is no hypothesis yet. */
    if (frame_idx < 0)
        return NO_BP;

    /* Now find the entry for </s> OR the best scoring entry. */
    assert(end_bpidx < ngs->bp_table_size);
    for (bp = ngs->bp_table_idx[frame_idx]; bp < end_bpidx; ++bp) {
        if (ngs->bp_table[bp].wid == ps_search_finish_wid(ngs)
            || ngs->bp_table[bp].score BETTER_THAN best_score) {
            best_score = ngs->bp_table[bp].score;
            best_exit = bp;
        }
        if (ngs->bp_table[bp].wid == ps_search_finish_wid(ngs))
            break;
    }

    if (out_best_score) {
	*out_best_score = best_score;
    }
    if (out_is_final) {
	*out_is_final = (ngs->bp_table[bp].wid == ps_search_finish_wid(ngs));
    }
    return best_exit;
}

char const *
ngram_search_bp_hyp(ngram_search_t *ngs, int bpidx)
{
    ps_search_t *base = ps_search_base(ngs);
    char *c;
    size_t len;
    int bp;

    if (bpidx == NO_BP)
        return NULL;

    bp = bpidx;
    len = 0;
    while (bp != NO_BP) {
        bptbl_t *be = &ngs->bp_table[bp];
        bp = be->bp;
        if (dict_real_word(ps_search_dict(ngs), be->wid))
            len += strlen(dict_basestr(ps_search_dict(ngs), be->wid)) + 1;
    }

    ckd_free(base->hyp_str);
    if (len == 0) {
	base->hyp_str = NULL;
	return base->hyp_str;
    }
    base->hyp_str = ckd_calloc(1, len);

    bp = bpidx;
    c = base->hyp_str + len - 1;
    while (bp != NO_BP) {
        bptbl_t *be = &ngs->bp_table[bp];
        size_t len;

        bp = be->bp;
        if (dict_real_word(ps_search_dict(ngs), be->wid)) {
            len = strlen(dict_basestr(ps_search_dict(ngs), be->wid));
            c -= len;
            memcpy(c, dict_basestr(ps_search_dict(ngs), be->wid), len);
            if (c > base->hyp_str) {
                --c;
                *c = ' ';
            }
        }
    }

    return base->hyp_str;
}

void
ngram_search_alloc_all_rc(ngram_search_t *ngs, int32 w)
{
    chan_t *hmm, *thmm;
    xwdssid_t *rssid;
    int32 i, tmatid, ciphone;

    /* DICT2PID */
    /* Get pointer to array of triphones for final diphone. */
    assert(!dict_is_single_phone(ps_search_dict(ngs), w));
    ciphone = dict_last_phone(ps_search_dict(ngs),w);
    rssid = dict2pid_rssid(ps_search_dict2pid(ngs),
                           ciphone,
                           dict_second_last_phone(ps_search_dict(ngs),w));
    tmatid = bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef, ciphone);
    hmm = ngs->word_chan[w];
    if ((hmm == NULL) || (hmm_nonmpx_ssid(&hmm->hmm) != rssid->ssid[0])) {
        hmm = listelem_malloc(ngs->chan_alloc);
        hmm->next = ngs->word_chan[w];
        ngs->word_chan[w] = hmm;

        hmm->info.rc_id = 0;
        hmm->ciphone = ciphone;
        hmm_init(ngs->hmmctx, &hmm->hmm, FALSE, rssid->ssid[0], tmatid);
        E_DEBUG(3,("allocated rc_id 0 ssid %d ciphone %d lc %d word %s\n",
                   rssid->ssid[0], hmm->ciphone,
                   dict_second_last_phone(ps_search_dict(ngs),w),
                   dict_wordstr(ps_search_dict(ngs),w)));
    }
    for (i = 1; i < rssid->n_ssid; ++i) {
        if ((hmm->next == NULL) || (hmm_nonmpx_ssid(&hmm->next->hmm) != rssid->ssid[i])) {
            thmm = listelem_malloc(ngs->chan_alloc);
            thmm->next = hmm->next;
            hmm->next = thmm;
            hmm = thmm;

            hmm->info.rc_id = i;
            hmm->ciphone = ciphone;
            hmm_init(ngs->hmmctx, &hmm->hmm, FALSE, rssid->ssid[i], tmatid);
            E_DEBUG(3,("allocated rc_id %d ssid %d ciphone %d lc %d word %s\n",
                       i, rssid->ssid[i], hmm->ciphone,
                       dict_second_last_phone(ps_search_dict(ngs),w),
                       dict_wordstr(ps_search_dict(ngs),w)));
        }
        else
            hmm = hmm->next;
    }
}

void
ngram_search_free_all_rc(ngram_search_t *ngs, int32 w)
{
    chan_t *hmm, *thmm;

    for (hmm = ngs->word_chan[w]; hmm; hmm = thmm) {
        thmm = hmm->next;
        hmm_deinit(&hmm->hmm);
        listelem_free(ngs->chan_alloc, hmm);
    }
    ngs->word_chan[w] = NULL;
}

int32
ngram_search_exit_score(ngram_search_t *ngs, bptbl_t *pbe, int rcphone)
{
    /* DICT2PID */
    /* Get the mapping from right context phone ID to index in the
     * right context table and the bscore_stack. */
    if (pbe->last2_phone == -1) {
        /* No right context for single phone predecessor words. */
        return pbe->score;
    }
    else {
        xwdssid_t *rssid;
        /* Find the index for the last diphone of the previous word +
         * the first phone of the current word. */
        rssid = dict2pid_rssid(ps_search_dict2pid(ngs),
                               pbe->last_phone, pbe->last2_phone);
        /* This may be WORST_SCORE, which means that there was no exit
         * with rcphone as right context. */
        return ngs->bscore_stack[pbe->s_idx + rssid->cimap[rcphone]];
    }
}

/*
 * Compute acoustic and LM scores for a BPTable entry (segment).
 */
void
ngram_compute_seg_score(ngram_search_t *ngs, bptbl_t *be, float32 lwf,
                        int32 *out_ascr, int32 *out_lscr)
{
    bptbl_t *pbe;
    int32 start_score;

    /* Start of utterance. */
    if (be->bp == NO_BP) {
        *out_ascr = be->score;
        *out_lscr = 0;
        return;
    }

    /* Otherwise, calculate lscr and ascr. */
    pbe = ngs->bp_table + be->bp;
    start_score = ngram_search_exit_score(ngs, pbe,
                                 dict_first_phone(ps_search_dict(ngs),be->wid));
    assert(start_score BETTER_THAN WORST_SCORE);

    /* FIXME: These result in positive acoustic scores when filler
       words have non-filler pronunciations.  That whole business
       is still pretty much broken but at least it doesn't
       segfault. */
    if (be->wid == ps_search_silence_wid(ngs)) {
        *out_lscr = ngs->silpen;
    }
    else if (dict_filler_word(ps_search_dict(ngs), be->wid)) {
        *out_lscr = ngs->fillpen;
    }
    else {
        int32 n_used;
        *out_lscr = ngram_tg_score(ngs->lmset,
                                   be->real_wid,
                                   pbe->real_wid,
                                   pbe->prev_real_wid,
                                   &n_used)>>SENSCR_SHIFT;
        *out_lscr = *out_lscr * lwf;
    }
    *out_ascr = be->score - start_score - *out_lscr;
}

static int
ngram_search_start(ps_search_t *search)
{
    ngram_search_t *ngs = (ngram_search_t *)search;

    ngs->done = FALSE;
    ngram_model_flush(ngs->lmset);
    if (ngs->fwdtree)
        ngram_fwdtree_start(ngs);
    else if (ngs->fwdflat)
        ngram_fwdflat_start(ngs);
    else
        return -1;
    return 0;
}

static int
ngram_search_step(ps_search_t *search, int frame_idx)
{
    ngram_search_t *ngs = (ngram_search_t *)search;

    if (ngs->fwdtree)
        return ngram_fwdtree_search(ngs, frame_idx);
    else if (ngs->fwdflat)
        return ngram_fwdflat_search(ngs, frame_idx);
    else
        return -1;
}

void
dump_bptable(ngram_search_t *ngs)
{
    int i;
    E_INFO("Backpointer table (%d entries):\n", ngs->bpidx);
    for (i = 0; i < ngs->bpidx; ++i) {
        bptbl_t *bpe = ngs->bp_table + i;
        int j, rcsize;

        E_INFO_NOFN("%-5d %-10s start %-3d end %-3d score %-8d bp %-3d real_wid %-5d prev_real_wid %-5d",
                    i, dict_wordstr(ps_search_dict(ngs), bpe->wid),
                    (bpe->bp == -1
                     ? 0 : ngs->bp_table[bpe->bp].frame + 1),
                    bpe->frame, bpe->score, bpe->bp,
                    bpe->real_wid, bpe->prev_real_wid);

        if (bpe->last2_phone == -1)
            rcsize = 0;
        else
            rcsize = dict2pid_rssid(ps_search_dict2pid(ngs),
                                    bpe->last_phone, bpe->last2_phone)->n_ssid;
        if (rcsize) {
            E_INFOCONT("\tbss");
            for (j = 0; j < rcsize; ++j)
                if (ngs->bscore_stack[bpe->s_idx + j] != WORST_SCORE)
                    E_INFOCONT(" %d", bpe->score - ngs->bscore_stack[bpe->s_idx + j]);
        }
        E_INFOCONT("\n");
    }
}

static int
ngram_search_finish(ps_search_t *search)
{
    ngram_search_t *ngs = (ngram_search_t *)search;

    ngs->n_tot_frame += ngs->n_frame;
    if (ngs->fwdtree) {
        ngram_fwdtree_finish(ngs);
        /* dump_bptable(ngs); */

        /* Now do fwdflat search in its entirety, if requested. */
        if (ngs->fwdflat) {
            int i;
            /* Rewind the acoustic model. */
            if (acmod_rewind(ps_search_acmod(ngs)) < 0)
                return -1;
            /* Now redo search. */
            ngram_fwdflat_start(ngs);
            i = 0;
            while (ps_search_acmod(ngs)->n_feat_frame > 0) {
                int nfr;
                if ((nfr = ngram_fwdflat_search(ngs, i)) < 0)
                    return nfr;
                acmod_advance(ps_search_acmod(ngs));
                ++i;
            }
            ngram_fwdflat_finish(ngs);
            /* And now, we should have a result... */
            /* dump_bptable(ngs); */
        }
    }
    else if (ngs->fwdflat) {
        ngram_fwdflat_finish(ngs);
    }

    /* Mark the current utterance as done. */
    ngs->done = TRUE;
    return 0;
}

static ps_latlink_t *
ngram_search_bestpath(ps_search_t *search, int32 *out_score, int backward)
{
    ngram_search_t *ngs = (ngram_search_t *)search;

    if (search->last_link == NULL) {
        search->last_link = ps_lattice_bestpath(search->dag, ngs->lmset,
                                                ngs->bestpath_fwdtree_lw_ratio,
                                                ngs->ascale);
        if (search->last_link == NULL)
            return NULL;
        /* Also calculate betas so we can fill in the posterior
         * probability field in the segmentation. */
        if (search->post == 0)
            search->post = ps_lattice_posterior(search->dag, ngs->lmset,
                                                ngs->ascale);
    }
    if (out_score)
        *out_score = search->last_link->path_scr + search->dag->final_node_ascr;
    return search->last_link;
}

static char const *
ngram_search_hyp(ps_search_t *search, int32 *out_score, int32 *out_is_final)
{
    ngram_search_t *ngs = (ngram_search_t *)search;

    /* Only do bestpath search if the utterance is complete. */
    if (ngs->bestpath && ngs->done) {
        ps_lattice_t *dag;
        ps_latlink_t *link;
        char const *hyp;
        double n_speech;

        ptmr_reset(&ngs->bestpath_perf);
        ptmr_start(&ngs->bestpath_perf);
        if ((dag = ngram_search_lattice(search)) == NULL)
            return NULL;
        if ((link = ngram_search_bestpath(search, out_score, FALSE)) == NULL)
            return NULL;
        hyp = ps_lattice_hyp(dag, link);
        ptmr_stop(&ngs->bestpath_perf);
        n_speech = (double)dag->n_frames
            / cmd_ln_int32_r(ps_search_config(ngs), "-frate");
        E_INFO("bestpath %.2f CPU %.3f xRT\n",
               ngs->bestpath_perf.t_cpu,
               ngs->bestpath_perf.t_cpu / n_speech);
        E_INFO("bestpath %.2f wall %.3f xRT\n",
               ngs->bestpath_perf.t_elapsed,
               ngs->bestpath_perf.t_elapsed / n_speech);
        return hyp;
    }
    else {
        int32 bpidx;

        /* fwdtree and fwdflat use same backpointer table. */
        bpidx = ngram_search_find_exit(ngs, -1, out_score, out_is_final);
        if (bpidx != NO_BP)
            return ngram_search_bp_hyp(ngs, bpidx);
    }

    return NULL;
}

static void
ngram_search_bp2itor(ps_seg_t *seg, int bp)
{
    ngram_search_t *ngs = (ngram_search_t *)seg->search;
    bptbl_t *be, *pbe;

    be = &ngs->bp_table[bp];
    pbe = be->bp == -1 ? NULL : &ngs->bp_table[be->bp];
    seg->word = dict_wordstr(ps_search_dict(ngs), be->wid);
    seg->ef = be->frame;
    seg->sf = pbe ? pbe->frame + 1 : 0;
    seg->prob = 0; /* Bogus value... */
    /* Compute acoustic and LM scores for this segment. */
    if (pbe == NULL) {
        seg->ascr = be->score;
        seg->lscr = 0;
        seg->lback = 0;
    }
    else {
        int32 start_score;

        /* Find ending path score of previous word. */
        start_score = ngram_search_exit_score(ngs, pbe,
                                     dict_first_phone(ps_search_dict(ngs), be->wid));
        assert(start_score BETTER_THAN WORST_SCORE);
        if (be->wid == ps_search_silence_wid(ngs)) {
            seg->lscr = ngs->silpen;
        }
        else if (dict_filler_word(ps_search_dict(ngs), be->wid)) {
            seg->lscr = ngs->fillpen;
        }
        else {
            seg->lscr = ngram_tg_score(ngs->lmset,
                                       be->real_wid,
                                       pbe->real_wid,
                                       pbe->prev_real_wid,
                                       &seg->lback)>>SENSCR_SHIFT;
            seg->lscr = (int32)(seg->lscr * seg->lwf);
        }
        seg->ascr = be->score - start_score - seg->lscr;
    }
}

static void
ngram_bp_seg_free(ps_seg_t *seg)
{
    bptbl_seg_t *itor = (bptbl_seg_t *)seg;

    ckd_free(itor->bpidx);
    ckd_free(itor);
}

static ps_seg_t *
ngram_bp_seg_next(ps_seg_t *seg)
{
    bptbl_seg_t *itor = (bptbl_seg_t *)seg;

    if (++itor->cur == itor->n_bpidx) {
        ngram_bp_seg_free(seg);
        return NULL;
    }

    ngram_search_bp2itor(seg, itor->bpidx[itor->cur]);
    return seg;
}

static ps_segfuncs_t ngram_bp_segfuncs = {
    /* seg_next */ ngram_bp_seg_next,
    /* seg_free */ ngram_bp_seg_free
};

static ps_seg_t *
ngram_search_bp_iter(ngram_search_t *ngs, int bpidx, float32 lwf)
{
    bptbl_seg_t *itor;
    int bp, cur;

    /* Calling this an "iterator" is a bit of a misnomer since we have
     * to get the entire backtrace in order to produce it.  On the
     * other hand, all we actually need is the bptbl IDs, and we can
     * allocate a fixed-size array of them. */
    itor = ckd_calloc(1, sizeof(*itor));
    itor->base.vt = &ngram_bp_segfuncs;
    itor->base.search = ps_search_base(ngs);
    itor->base.lwf = lwf;
    itor->n_bpidx = 0;
    bp = bpidx;
    while (bp != NO_BP) {
        bptbl_t *be = &ngs->bp_table[bp];
        bp = be->bp;
        ++itor->n_bpidx;
    }
    if (itor->n_bpidx == 0) {
        ckd_free(itor);
        return NULL;
    }
    itor->bpidx = ckd_calloc(itor->n_bpidx, sizeof(*itor->bpidx));
    cur = itor->n_bpidx - 1;
    bp = bpidx;
    while (bp != NO_BP) {
        bptbl_t *be = &ngs->bp_table[bp];
        itor->bpidx[cur] = bp;
        bp = be->bp;
        --cur;
    }

    /* Fill in relevant fields for first element. */
    ngram_search_bp2itor((ps_seg_t *)itor, itor->bpidx[0]);

    return (ps_seg_t *)itor;
}

static ps_seg_t *
ngram_search_seg_iter(ps_search_t *search, int32 *out_score)
{
    ngram_search_t *ngs = (ngram_search_t *)search;

    /* Only do bestpath search if the utterance is done. */
    if (ngs->bestpath && ngs->done) {
        ps_lattice_t *dag;
        ps_latlink_t *link;
        double n_speech;
        ps_seg_t *itor;

        ptmr_reset(&ngs->bestpath_perf);
        ptmr_start(&ngs->bestpath_perf);
        if ((dag = ngram_search_lattice(search)) == NULL)
            return NULL;
        if ((link = ngram_search_bestpath(search, out_score, TRUE)) == NULL)
            return NULL;
        itor = ps_lattice_seg_iter(dag, link,
                                   ngs->bestpath_fwdtree_lw_ratio);
        ptmr_stop(&ngs->bestpath_perf);
        n_speech = (double)dag->n_frames
            / cmd_ln_int32_r(ps_search_config(ngs), "-frate");
        E_INFO("bestpath %.2f CPU %.3f xRT\n",
               ngs->bestpath_perf.t_cpu,
               ngs->bestpath_perf.t_cpu / n_speech);
        E_INFO("bestpath %.2f wall %.3f xRT\n",
               ngs->bestpath_perf.t_elapsed,
               ngs->bestpath_perf.t_elapsed / n_speech);
        return itor;
    }
    else {
        int32 bpidx;

        /* fwdtree and fwdflat use same backpointer table. */
        bpidx = ngram_search_find_exit(ngs, -1, out_score, NULL);
        return ngram_search_bp_iter(ngs, bpidx,
                                    /* but different language weights... */
                                    (ngs->done && ngs->fwdflat)
                                    ? ngs->fwdflat_fwdtree_lw_ratio : 1.0);
    }

    return NULL;
}

static int32
ngram_search_prob(ps_search_t *search)
{
    ngram_search_t *ngs = (ngram_search_t *)search;

    /* Only do bestpath search if the utterance is done. */
    if (ngs->bestpath && ngs->done) {
        ps_lattice_t *dag;
        ps_latlink_t *link;

        if ((dag = ngram_search_lattice(search)) == NULL)
            return 0;
        if ((link = ngram_search_bestpath(search, NULL, TRUE)) == NULL)
            return 0;
        return search->post;
    }
    else {
        /* FIXME: Give some kind of good estimate here, eventually. */
        return 0;
    }
}

static void
create_dag_nodes(ngram_search_t *ngs, ps_lattice_t *dag)
{
    bptbl_t *bp_ptr;
    int32 i;

    for (i = 0, bp_ptr = ngs->bp_table; i < ngs->bpidx; ++i, ++bp_ptr) {
        int32 sf, ef, wid;
        ps_latnode_t *node;

        /* Skip invalid backpointers (these result from -maxwpf pruning) */
        if (!bp_ptr->valid)
            continue;

        sf = (bp_ptr->bp < 0) ? 0 : ngs->bp_table[bp_ptr->bp].frame + 1;
        ef = bp_ptr->frame;
        wid = bp_ptr->wid;

        assert(ef < dag->n_frames);
        /* Skip non-final </s> entries. */
        if ((wid == ps_search_finish_wid(ngs)) && (ef < dag->n_frames - 1))
            continue;

        /* Skip if word not in LM */
        if ((!dict_filler_word(ps_search_dict(ngs), wid))
            && (!ngram_model_set_known_wid(ngs->lmset,
                                           dict_basewid(ps_search_dict(ngs), wid))))
            continue;

        /* See if bptbl entry <wid,sf> already in lattice */
        for (node = dag->nodes; node; node = node->next) {
            if ((node->wid == wid) && (node->sf == sf))
                break;
        }

        /* For the moment, store bptbl indices in node.{fef,lef} */
        if (node)
            node->lef = i;
        else {
            /* New node; link to head of list */
            node = listelem_malloc(dag->latnode_alloc);
            node->wid = wid;
            node->sf = sf; /* This is a frame index. */
            node->fef = node->lef = i; /* These are backpointer indices (argh) */
            node->reachable = FALSE;
            node->entries = NULL;
            node->exits = NULL;

            /* NOTE: This creates the list of nodes in reverse
             * topological order, i.e. a node always precedes its
             * antecedents in this list. */
            node->next = dag->nodes;
            dag->nodes = node;
            ++dag->n_nodes;
        }
    }
}

static ps_latnode_t *
find_start_node(ngram_search_t *ngs, ps_lattice_t *dag)
{
    ps_latnode_t *node;

    /* Find start node <s>.0 */
    for (node = dag->nodes; node; node = node->next) {
        if ((node->wid == ps_search_start_wid(ngs)) && (node->sf == 0))
            break;
    }
    if (!node) {
        /* This is probably impossible. */
        E_ERROR("Couldn't find <s> in first frame\n");
        return NULL;
    }
    return node;
}

static ps_latnode_t *
find_end_node(ngram_search_t *ngs, ps_lattice_t *dag, float32 lwf)
{
    ps_latnode_t *node;
    int32 ef, bestbp, bp, bestscore;

    /* Find final node </s>.last_frame; nothing can follow this node */
    for (node = dag->nodes; node; node = node->next) {
        int32 lef = ngs->bp_table[node->lef].frame;
        if ((node->wid == ps_search_finish_wid(ngs))
            && (lef == dag->n_frames - 1))
            break;
    }
    if (node != NULL)
        return node;

    /* It is quite likely that no </s> exited in the last frame.  So,
     * find the node corresponding to the best exit. */
    /* Find the last frame containing a word exit. */
    for (ef = dag->n_frames - 1;
         ef >= 0 && ngs->bp_table_idx[ef] == ngs->bpidx;
         --ef);
    if (ef < 0) {
        E_ERROR("Empty backpointer table: can not build DAG.\n");
        return NULL;
    }

    /* Find best word exit in that frame. */
    bestscore = WORST_SCORE;
    bestbp = NO_BP;
    for (bp = ngs->bp_table_idx[ef]; bp < ngs->bp_table_idx[ef + 1]; ++bp) {
        int32 n_used, l_scr, wid, prev_wid;
        wid = ngs->bp_table[bp].real_wid;
        prev_wid = ngs->bp_table[bp].prev_real_wid;
        /* Always prefer </s>, of which there will only be one per frame. */
        if (wid == ps_search_finish_wid(ngs)) {
            bestbp = bp;
            break;
        }
        l_scr = ngram_tg_score(ngs->lmset, ps_search_finish_wid(ngs),
                               wid, prev_wid, &n_used) >>SENSCR_SHIFT;
        l_scr = l_scr * lwf;
        if (ngs->bp_table[bp].score + l_scr BETTER_THAN bestscore) {
            bestscore = ngs->bp_table[bp].score + l_scr;
            bestbp = bp;
        }
    }
    if (bestbp == NO_BP) {
        E_ERROR("No word exits found in last frame (%d), assuming no recognition\n", ef);
        return NULL;
    }
    E_INFO("</s> not found in last frame, using %s.%d instead\n",
           dict_basestr(ps_search_dict(ngs), ngs->bp_table[bestbp].wid), ef);

    /* Now find the node that corresponds to it. */
    for (node = dag->nodes; node; node = node->next) {
        if (node->lef == bestbp)
            return node;
    }

    /* FIXME: This seems to happen a lot! */
    E_ERROR("Failed to find DAG node corresponding to %s\n",
           dict_basestr(ps_search_dict(ngs), ngs->bp_table[bestbp].wid));
    return NULL;
}

/*
 * Build lattice from bptable.
 */
ps_lattice_t *
ngram_search_lattice(ps_search_t *search)
{
    int32 i, score, ascr, lscr;
    ps_latnode_t *node, *from, *to;
    ngram_search_t *ngs;
    ps_lattice_t *dag;
    int min_endfr, nlink;
    float lwf;

    ngs = (ngram_search_t *)search;
    min_endfr = cmd_ln_int32_r(ps_search_config(search), "-min_endfr");

    /* If the best score is WORST_SCORE or worse, there is no way to
     * make a lattice. */
    if (ngs->best_score == WORST_SCORE || ngs->best_score WORSE_THAN WORST_SCORE)
        return NULL;

    /* Check to see if a lattice has previously been created over the
     * same number of frames, and reuse it if so. */
    if (search->dag && search->dag->n_frames == ngs->n_frame)
        return search->dag;

    /* Nope, create a new one. */
    ps_lattice_free(search->dag);
    search->dag = NULL;
    dag = ps_lattice_init_search(search, ngs->n_frame);
    /* Compute these such that they agree with the fwdtree language weight. */
    lwf = ngs->fwdflat ? ngs->fwdflat_fwdtree_lw_ratio : 1.0;
    create_dag_nodes(ngs, dag);
    if ((dag->start = find_start_node(ngs, dag)) == NULL)
        goto error_out;
    if ((dag->end = find_end_node(ngs, dag, ngs->bestpath_fwdtree_lw_ratio)) == NULL)
        goto error_out;
    E_INFO("lattice start node %s.%d end node %s.%d\n",
           dict_wordstr(search->dict, dag->start->wid), dag->start->sf,
           dict_wordstr(search->dict, dag->end->wid), dag->end->sf);

    ngram_compute_seg_score(ngs, ngs->bp_table + dag->end->lef, lwf,
                            &dag->final_node_ascr, &lscr);

    /*
     * At this point, dag->nodes is ordered such that nodes earlier in
     * the list can follow (in time) those later in the list, but not
     * vice versa (see above - also note that adjacency is purely
     * determined by time which is why we can make this claim).  Now
     * create precedence links and simultanesously mark all nodes that
     * can reach dag->end.  (All nodes are reached from dag->start
     * simply by definition - they were created that way).
     *
     * Note that this also means that any nodes before dag->end in the
     * list can be discarded, meaning that dag->end will always be
     * equal to dag->nodes (FIXME: except when loading from a file but
     * we can fix that...)
     */
    i = 0;
    while (dag->nodes && dag->nodes != dag->end) {
        ps_latnode_t *next = dag->nodes->next;
        listelem_free(dag->latnode_alloc, dag->nodes);
        dag->nodes = next;
        ++i;
    }
    E_INFO("Eliminated %d nodes before end node\n", i);
    dag->end->reachable = TRUE;
    nlink = 0;
    for (to = dag->end; to; to = to->next) {
        int fef, lef;

        /* Skip if not reachable; it will never be reachable from dag->end */
        if (!to->reachable)
            continue;

        /* Prune nodes with too few endpoints - heuristic
           borrowed from Sphinx3 */
        fef = ngs->bp_table[to->fef].frame;
        lef = ngs->bp_table[to->lef].frame;
        if (to != dag->end && lef - fef < min_endfr) {
            to->reachable = FALSE;
            continue;
        }

        /* Find predecessors of to : from->fef+1 <= to->sf <= from->lef+1 */
        for (from = to->next; from; from = from->next) {
            bptbl_t *from_bpe;

            fef = ngs->bp_table[from->fef].frame;
            lef = ngs->bp_table[from->lef].frame;

            if ((to->sf <= fef) || (to->sf > lef + 1))
                continue;
            if (lef - fef < min_endfr) {
                assert(!from->reachable);
                continue;
            }

            /* Find bptable entry for "from" that exactly precedes "to" */
            i = from->fef;
            from_bpe = ngs->bp_table + i;
            for (; i <= from->lef; i++, from_bpe++) {
                if (from_bpe->wid != from->wid)
                    continue;
                if (from_bpe->frame >= to->sf - 1)
                    break;
            }

            if ((i > from->lef) || (from_bpe->frame != to->sf - 1))
                continue;

            /* Find acoustic score from.sf->to.sf-1 with right context = to */
            /* This gives us from_bpe's best acoustic score. */
            ngram_compute_seg_score(ngs, from_bpe, lwf,
                                    &ascr, &lscr);
            /* Now find the exact path score for from->to, including
             * the appropriate final triphone.  In fact this might not
             * exist. */
            score = ngram_search_exit_score(ngs, from_bpe,
                                            dict_first_phone(ps_search_dict(ngs), to->wid));
            /* Does not exist.  Can't create a link here. */
            if (score == WORST_SCORE)
                continue;
            /* Adjust the arc score to match the correct triphone. */
            else
                score = ascr + (score - from_bpe->score);
            if (score BETTER_THAN 0) {
                /* Scores must be negative, or Bad Things will happen.
                   In general, they are, except in corner cases
                   involving filler words.  We don't want to throw any
                   links away so we'll keep these, but with some
                   arbitrarily improbable but recognizable score. */
                ps_lattice_link(dag, from, to, -424242, from_bpe->frame);
                ++nlink;
                from->reachable = TRUE;
            }
            else if (score BETTER_THAN WORST_SCORE) {
                ps_lattice_link(dag, from, to, score, from_bpe->frame);
                ++nlink;
                from->reachable = TRUE;
            }
        }
    }

    /* There must be at least one path between dag->start and dag->end */
    if (!dag->start->reachable) {
        E_ERROR("End node of lattice isolated; unreachable\n");
        goto error_out;
    }

    for (node = dag->nodes; node; node = node->next) {
        /* Change node->{fef,lef} from bptbl indices to frames. */
        node->fef = ngs->bp_table[node->fef].frame;
        node->lef = ngs->bp_table[node->lef].frame;
        /* Find base wid for nodes. */
        node->basewid = dict_basewid(search->dict, node->wid);
    }

    /* Link nodes with alternate pronunciations at the same timepoint. */
    for (node = dag->nodes; node; node = node->next) {
        ps_latnode_t *alt;
        /* Scan forward to find the next alternate, then stop. */
        for (alt = node->next; alt && alt->sf == node->sf; alt = alt->next) {
            if (alt->basewid == node->basewid) {
                alt->alt = node->alt;
                node->alt = alt;
                break;
            }
        }
    }
    E_INFO("Lattice has %d nodes, %d links\n", dag->n_nodes, nlink);

    /* Minor hack: If the final node is a filler word and not </s>,
     * then set its base word ID to </s>, so that the language model
     * scores won't be screwed up. */
    if (dict_filler_word(ps_search_dict(ngs), dag->end->wid))
        dag->end->basewid = ps_search_finish_wid(ngs);

    /* Free nodes unreachable from dag->end and their links */
    ps_lattice_delete_unreachable(dag);

    /* Add silprob and fillprob to corresponding links */
    ps_lattice_penalize_fillers(dag, ngs->silpen, ngs->fillpen);

    search->dag = dag;
    return dag;

error_out:
    ps_lattice_free(dag);
    return NULL;
}

void ngram_search_set_lm(ngram_model_t *lm)
{
    default_lm = ngram_model_retain(lm);
}