summaryrefslogtreecommitdiffstats
path: root/media/sphinxbase/src/libsphinxbase/lm/ngram_model_arpa.c
blob: a4b72cb00527c3e540335551ce04bb5afa4a20fe (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
/* -*- c-basic-offset: 4; indent-tabs-mode: nil -*- */
/* ====================================================================
 * Copyright (c) 1999-2007 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_model_arpa.c ARPA format language models
 *
 * Author: David Huggins-Daines <dhuggins@cs.cmu.edu>
 */

#include "sphinxbase/ckd_alloc.h"
#include <string.h>
#include <limits.h>
#include <assert.h>

#include "sphinxbase/err.h"
#include "sphinxbase/pio.h"
#include "sphinxbase/listelem_alloc.h"
#include "sphinxbase/strfuncs.h"

#include "ngram_model_arpa.h"

static ngram_funcs_t ngram_model_arpa_funcs;

#define TSEG_BASE(m,b)		((m)->lm3g.tseg_base[(b)>>LOG_BG_SEG_SZ])
#define FIRST_BG(m,u)		((m)->lm3g.unigrams[u].bigrams)
#define FIRST_TG(m,b)		(TSEG_BASE((m),(b))+((m)->lm3g.bigrams[b].trigrams))

/*
 * Read and return #unigrams, #bigrams, #trigrams as stated in input file.
 */
static int
ReadNgramCounts(lineiter_t **li, int32 * n_ug, int32 * n_bg, int32 * n_tg)
{
    int32 ngram, ngram_cnt;

    /* skip file until past the '\data\' marker */
    while (*li) {
        string_trim((*li)->buf, STRING_BOTH);
        if (strcmp((*li)->buf, "\\data\\") == 0)
            break;
        *li = lineiter_next(*li);
    }
    if (*li == NULL || strcmp((*li)->buf, "\\data\\") != 0) {
        E_INFO("No \\data\\ mark in LM file\n");
        return -1;
    }

    *n_ug = *n_bg = *n_tg = 0;
    while ((*li = lineiter_next(*li))) {
        if (sscanf((*li)->buf, "ngram %d=%d", &ngram, &ngram_cnt) != 2)
            break;
        switch (ngram) {
        case 1:
            *n_ug = ngram_cnt;
            break;
        case 2:
            *n_bg = ngram_cnt;
            break;
        case 3:
            *n_tg = ngram_cnt;
            break;
        default:
            E_ERROR("Unknown ngram (%d)\n", ngram);
            return -1;
        }
    }
    if (*li == NULL) {
        E_ERROR("EOF while reading ngram counts\n");
        return -1;
    }

    /* Position iterator to the unigrams header '\1-grams:\' */
    while ((*li = lineiter_next(*li))) {
        string_trim((*li)->buf, STRING_BOTH);
        if (strcmp((*li)->buf, "\\1-grams:") == 0)
            break;
    }
    if (*li == NULL) {
        E_ERROR_SYSTEM("Failed to read \\1-grams: mark");
        return -1;
    }

    if ((*n_ug <= 0) || (*n_bg < 0) || (*n_tg < 0)) {
        E_ERROR("Bad or missing ngram count\n");
        return -1;
    }
    return 0;
}

/*
 * Read in the unigrams from given file into the LM structure model.
 * On entry to this procedure, the iterator is positioned to the
 * header line '\1-grams:'.
 */
static int
ReadUnigrams(lineiter_t **li, ngram_model_arpa_t * model)
{
    ngram_model_t *base = &model->base;
    int32 wcnt;
    float p1;

    E_INFO("Reading unigrams\n");

    wcnt = 0;
    while ((*li = lineiter_next(*li))) {
        char *wptr[3], *name;
        float32 bo_wt = 0.0f;
        int n;

        string_trim((*li)->buf, STRING_BOTH);
        if (strcmp((*li)->buf, "\\2-grams:") == 0
            || strcmp((*li)->buf, "\\end\\") == 0)
            break;

        if ((n = str2words((*li)->buf, wptr, 3)) < 2) {
            if ((*li)->buf[0] != '\0')
                E_WARN("Format error; unigram ignored: %s\n", (*li)->buf);
            continue;
        }
        else {
            p1 = (float)atof_c(wptr[0]);
            name = wptr[1];
            if (n == 3)
                bo_wt = (float)atof_c(wptr[2]);
        }

        if (wcnt >= base->n_counts[0]) {
            E_ERROR("Too many unigrams\n");
            return -1;
        }

        /* Associate name with word id */
        base->word_str[wcnt] = ckd_salloc(name);
        if ((hash_table_enter(base->wid, base->word_str[wcnt], (void *)(long)wcnt))
            != (void *)(long)wcnt) {
                E_WARN("Duplicate word in dictionary: %s\n", base->word_str[wcnt]);
        }
        model->lm3g.unigrams[wcnt].prob1.l = logmath_log10_to_log(base->lmath, p1);
        model->lm3g.unigrams[wcnt].bo_wt1.l = logmath_log10_to_log(base->lmath, bo_wt);
        wcnt++;
    }

    if (base->n_counts[0] != wcnt) {
        E_WARN("lm_t.ucount(%d) != #unigrams read(%d)\n",
               base->n_counts[0], wcnt);
        base->n_counts[0] = wcnt;
        base->n_words = wcnt;
    }
    return 0;
}

/*
 * Read bigrams from given file into given model structure.
 */
static int
ReadBigrams(lineiter_t **li, ngram_model_arpa_t * model)
{
    ngram_model_t *base = &model->base;
    int32 w1, w2, prev_w1, bgcount;
    bigram_t *bgptr;

    E_INFO("Reading bigrams\n");

    bgcount = 0;
    bgptr = model->lm3g.bigrams;
    prev_w1 = -1;

    while ((*li = lineiter_next(*li))) {
        float32 p, bo_wt = 0.0f;
        int32 p2, bo_wt2;
        char *wptr[4], *word1, *word2;
        int n;

        string_trim((*li)->buf, STRING_BOTH);
        wptr[3] = NULL;
        if ((n = str2words((*li)->buf, wptr, 4)) < 3) {
            if ((*li)->buf[0] != '\0')
                break;
            continue;
        }
        else {
            p = (float32)atof_c(wptr[0]);
            word1 = wptr[1];
            word2 = wptr[2];
            if (wptr[3])
                bo_wt = (float32)atof_c(wptr[3]);
        }

        if ((w1 = ngram_wid(base, word1)) == NGRAM_INVALID_WID) {
            E_ERROR("Unknown word: %s, skipping bigram (%s %s)\n",
                    word1, word1, word2);
            continue;
        }
        if ((w2 = ngram_wid(base, word2)) == NGRAM_INVALID_WID) {
            E_ERROR("Unknown word: %s, skipping bigram (%s %s)\n",
                    word2, word1, word2);
            continue;
        }

        /* FIXME: Should use logmath_t quantization here. */
        /* HACK!! to quantize probs to 4 decimal digits */
        p = (float32)((int32)(p * 10000)) / 10000;
        bo_wt = (float32)((int32)(bo_wt * 10000)) / 10000;

        p2 = logmath_log10_to_log(base->lmath, p);
        bo_wt2 = logmath_log10_to_log(base->lmath, bo_wt);

        if (bgcount >= base->n_counts[1]) {
            E_ERROR("Too many bigrams\n");
            return -1;
        }

        bgptr->wid = w2;
        bgptr->prob2 = sorted_id(&model->sorted_prob2, &p2);
        if (base->n_counts[2] > 0)
            bgptr->bo_wt2 = sorted_id(&model->sorted_bo_wt2, &bo_wt2);

        if (w1 != prev_w1) {
            if (w1 < prev_w1) {
                E_ERROR("Bigram %s %s not in unigram order word id: %d prev word id: %d\n", word1, word2, w1, prev_w1);
                return -1;
            }

            for (prev_w1++; prev_w1 <= w1; prev_w1++)
                model->lm3g.unigrams[prev_w1].bigrams = bgcount;
            prev_w1 = w1;
        }
        bgcount++;
        bgptr++;

        if ((bgcount & 0x0000ffff) == 0) {
            E_INFOCONT(".");
        }
    }
    if (*li == NULL || ((strcmp((*li)->buf, "\\end\\") != 0)
                        && (strcmp((*li)->buf, "\\3-grams:") != 0))) {
        E_ERROR("Bad bigram: %s\n", (*li)->buf);
        return -1;
    }

    for (prev_w1++; prev_w1 <= base->n_counts[0]; prev_w1++)
        model->lm3g.unigrams[prev_w1].bigrams = bgcount;

    return 0;
}

/*
 * Very similar to ReadBigrams.
 */
static int
ReadTrigrams(lineiter_t **li, ngram_model_arpa_t * model)
{
    ngram_model_t *base = &model->base;
    int32 i, w1, w2, w3, prev_w1, prev_w2, tgcount, prev_bg, bg, endbg;
    int32 seg, prev_seg, prev_seg_lastbg;
    trigram_t *tgptr;
    bigram_t *bgptr;

    E_INFO("Reading trigrams\n");

    tgcount = 0;
    tgptr = model->lm3g.trigrams;
    prev_w1 = -1;
    prev_w2 = -1;
    prev_bg = -1;
    prev_seg = -1;

    while ((*li = lineiter_next(*li))) {
        float32 p;
        int32 p3;
        char *wptr[4], *word1, *word2, *word3;

        string_trim((*li)->buf, STRING_BOTH);
        if (str2words((*li)->buf, wptr, 4) != 4) {
            if ((*li)->buf[0] != '\0')
                break;
            continue;
        }
        else {
            p = (float32)atof_c(wptr[0]);
            word1 = wptr[1];
            word2 = wptr[2];
            word3 = wptr[3];
        }

        if ((w1 = ngram_wid(base, word1)) == NGRAM_INVALID_WID) {
            E_ERROR("Unknown word: %s, skipping trigram (%s %s %s)\n",
                    word1, word1, word2, word3);
            continue;
        }
        if ((w2 = ngram_wid(base, word2)) == NGRAM_INVALID_WID) {
            E_ERROR("Unknown word: %s, skipping trigram (%s %s %s)\n",
                    word2, word1, word2, word3);
            continue;
        }
        if ((w3 = ngram_wid(base, word3)) == NGRAM_INVALID_WID) {
            E_ERROR("Unknown word: %s, skipping trigram (%s %s %s)\n",
                    word3, word1, word2, word3);
            continue;
        }

        /* FIXME: Should use logmath_t quantization here. */
        /* HACK!! to quantize probs to 4 decimal digits */
        p = (float32)((int32)(p * 10000)) / 10000;
        p3 = logmath_log10_to_log(base->lmath, p);

        if (tgcount >= base->n_counts[2]) {
            E_ERROR("Too many trigrams\n");
            return -1;
        }

        tgptr->wid = w3;
        tgptr->prob3 = sorted_id(&model->sorted_prob3, &p3);

        if ((w1 != prev_w1) || (w2 != prev_w2)) {
            /* Trigram for a new bigram; update tg info for all previous bigrams */
            if ((w1 < prev_w1) || ((w1 == prev_w1) && (w2 < prev_w2))) {
                E_ERROR("Trigrams not in bigram order\n");
                return -1;
            }

            bg = (w1 !=
                  prev_w1) ? model->lm3g.unigrams[w1].bigrams : prev_bg + 1;
            endbg = model->lm3g.unigrams[w1 + 1].bigrams;
            bgptr = model->lm3g.bigrams + bg;
            for (; (bg < endbg) && (bgptr->wid != w2); bg++, bgptr++);
            if (bg >= endbg) {
                E_ERROR("Missing bigram for trigram: %s", (*li)->buf);
                return -1;
            }

            /* bg = bigram entry index for <w1,w2>.  Update tseg_base */
            seg = bg >> LOG_BG_SEG_SZ;
            for (i = prev_seg + 1; i <= seg; i++)
                model->lm3g.tseg_base[i] = tgcount;

            /* Update trigrams pointers for all bigrams until bg */
            if (prev_seg < seg) {
                int32 tgoff = 0;

                if (prev_seg >= 0) {
                    tgoff = tgcount - model->lm3g.tseg_base[prev_seg];
                    if (tgoff > 65535) {
                        E_ERROR("Size of trigram segment is bigger than 65535, such a big language models are not supported, use smaller vocabulary\n");
                        return -1;
                    }
                }

                prev_seg_lastbg = ((prev_seg + 1) << LOG_BG_SEG_SZ) - 1;
                bgptr = model->lm3g.bigrams + prev_bg;
                for (++prev_bg, ++bgptr; prev_bg <= prev_seg_lastbg;
                     prev_bg++, bgptr++)
                    bgptr->trigrams = tgoff;

                for (; prev_bg <= bg; prev_bg++, bgptr++)
                    bgptr->trigrams = 0;
            }
            else {
                int32 tgoff;

                tgoff = tgcount - model->lm3g.tseg_base[prev_seg];
                if (tgoff > 65535) {
                    E_ERROR("Size of trigram segment is bigger than 65535, such a big language models are not supported, use smaller vocabulary\n");
                    return -1;
                }

                bgptr = model->lm3g.bigrams + prev_bg;
                for (++prev_bg, ++bgptr; prev_bg <= bg; prev_bg++, bgptr++)
                    bgptr->trigrams = tgoff;
            }

            prev_w1 = w1;
            prev_w2 = w2;
            prev_bg = bg;
            prev_seg = seg;
        }

        tgcount++;
        tgptr++;

        if ((tgcount & 0x0000ffff) == 0) {
            E_INFOCONT(".");
        }
    }
    if (*li == NULL || strcmp((*li)->buf, "\\end\\") != 0) {
        E_ERROR("Bad trigram: %s\n", (*li)->buf);
        return -1;
    }

    for (prev_bg++; prev_bg <= base->n_counts[1]; prev_bg++) {
        if ((prev_bg & (BG_SEG_SZ - 1)) == 0)
            model->lm3g.tseg_base[prev_bg >> LOG_BG_SEG_SZ] = tgcount;
        if ((tgcount - model->lm3g.tseg_base[prev_bg >> LOG_BG_SEG_SZ]) > 65535) {
            E_ERROR("Size of trigram segment is bigger than 65535, such a big language models are not supported, use smaller vocabulary\n");
            return -1;
        }
        model->lm3g.bigrams[prev_bg].trigrams =
            tgcount - model->lm3g.tseg_base[prev_bg >> LOG_BG_SEG_SZ];
    }
    return 0;
}

static unigram_t *
new_unigram_table(int32 n_ug)
{
    unigram_t *table;
    int32 i;

    table = ckd_calloc(n_ug, sizeof(unigram_t));
    for (i = 0; i < n_ug; i++) {
        table[i].prob1.l = INT_MIN;
        table[i].bo_wt1.l = INT_MIN;
    }
    return table;
}

ngram_model_t *
ngram_model_arpa_read(cmd_ln_t *config,
		      const char *file_name,
		      logmath_t *lmath)
{
    lineiter_t *li;
    FILE *fp;
    int32 is_pipe;
    int32 n_unigram;
    int32 n_bigram;
    int32 n_trigram;
    int32 n;
    ngram_model_arpa_t *model;
    ngram_model_t *base;

    if ((fp = fopen_comp(file_name, "r", &is_pipe)) == NULL) {
        E_ERROR("File %s not found\n", file_name);
        return NULL;
    }
    li = lineiter_start(fp);
 
    /* Read #unigrams, #bigrams, #trigrams from file */
    if (ReadNgramCounts(&li, &n_unigram, &n_bigram, &n_trigram) == -1) {
        lineiter_free(li);
        fclose_comp(fp, is_pipe);
        return NULL;
    }
    E_INFO("ngrams 1=%d, 2=%d, 3=%d\n", n_unigram, n_bigram, n_trigram);

    /* Allocate space for LM, including initial OOVs and placeholders; initialize it */
    model = ckd_calloc(1, sizeof(*model));
    base = &model->base;
    if (n_trigram > 0)
        n = 3;
    else if (n_bigram > 0)
        n = 2;
    else
        n = 1;
    /* Initialize base model. */
    ngram_model_init(base, &ngram_model_arpa_funcs, lmath, n, n_unigram);
    base->n_counts[0] = n_unigram;
    base->n_counts[1] = n_bigram;
    base->n_counts[2] = n_trigram;
    base->writable = TRUE;

    /*
     * Allocate one extra unigram and bigram entry: sentinels to terminate
     * followers (bigrams and trigrams, respectively) of previous entry.
     */
    model->lm3g.unigrams = new_unigram_table(n_unigram + 1);
    model->lm3g.bigrams =
        ckd_calloc(n_bigram + 1, sizeof(bigram_t));
    if (n_trigram > 0)
        model->lm3g.trigrams =
            ckd_calloc(n_trigram, sizeof(trigram_t));

    if (n_trigram > 0) {
        model->lm3g.tseg_base =
            ckd_calloc((n_bigram + 1) / BG_SEG_SZ + 1,
                       sizeof(int32));
    }
    if (ReadUnigrams(&li, model) == -1) {
        fclose_comp(fp, is_pipe);
        ngram_model_free(base);
        return NULL;
    }
    E_INFO("%8d = #unigrams created\n", base->n_counts[0]);

    if (base->n_counts[2] > 0)
        init_sorted_list(&model->sorted_bo_wt2);

    if (base->n_counts[1] > 0) {
        init_sorted_list(&model->sorted_prob2);

        if (ReadBigrams(&li, model) == -1) {
            fclose_comp(fp, is_pipe);
            ngram_model_free(base);
            return NULL;
        }

        base->n_counts[1] = FIRST_BG(model, base->n_counts[0]);
        model->lm3g.n_prob2 = model->sorted_prob2.free;
        model->lm3g.prob2 = vals_in_sorted_list(&model->sorted_prob2);
        free_sorted_list(&model->sorted_prob2);
        E_INFO("%8d = #bigrams created\n", base->n_counts[1]);
        E_INFO("%8d = #prob2 entries\n", model->lm3g.n_prob2);
    }

    if (base->n_counts[2] > 0) {
        /* Create trigram bo-wts array */
        model->lm3g.n_bo_wt2 = model->sorted_bo_wt2.free;
        model->lm3g.bo_wt2 = vals_in_sorted_list(&model->sorted_bo_wt2);
        free_sorted_list(&model->sorted_bo_wt2);
        E_INFO("%8d = #bo_wt2 entries\n", model->lm3g.n_bo_wt2);

        init_sorted_list(&model->sorted_prob3);

        if (ReadTrigrams(&li, model) == -1) {
            fclose_comp(fp, is_pipe);
            ngram_model_free(base);
            return NULL;
        }

        base->n_counts[2] = FIRST_TG(model, base->n_counts[1]);
        model->lm3g.n_prob3 = model->sorted_prob3.free;
        model->lm3g.prob3 = vals_in_sorted_list(&model->sorted_prob3);
        E_INFO("%8d = #trigrams created\n", base->n_counts[2]);
        E_INFO("%8d = #prob3 entries\n", model->lm3g.n_prob3);

        free_sorted_list(&model->sorted_prob3);

        /* Initialize tginfo */
        model->lm3g.tginfo = ckd_calloc(n_unigram, sizeof(tginfo_t *));
        model->lm3g.le = listelem_alloc_init(sizeof(tginfo_t));
    }

    lineiter_free(li);
    fclose_comp(fp, is_pipe);
    return base;
}

int
ngram_model_arpa_write(ngram_model_t *model,
		       const char *file_name)
{
    ngram_iter_t *itor;
    FILE *fh;
    int i;

    if ((fh = fopen(file_name, "w")) == NULL) {
        E_ERROR_SYSTEM("Failed to open %s for writing", file_name);
        return -1;
    }
    fprintf(fh, "This is an ARPA-format language model file, generated by CMU Sphinx\n");

    /* The ARPA format doesn't require any extra information that
     * N-Gram iterators can't give us, so this is very
     * straightforward compared with DMP writing. */

    /* Write N-gram counts. */
    fprintf(fh, "\\data\\\n");
    for (i = 0; i < model->n; ++i) {
        fprintf(fh, "ngram %d=%d\n", i+1, model->n_counts[i]);
    }

    /* Write N-grams */
    for (i = 0; i < model->n; ++i) {
        fprintf(fh, "\n\\%d-grams:\n", i + 1);
        for (itor = ngram_model_mgrams(model, i); itor; itor = ngram_iter_next(itor)) {
            int32 const *wids;
            int32 score, bowt;
            int j;

            wids = ngram_iter_get(itor, &score, &bowt);
            fprintf(fh, "%.4f ", logmath_log_to_log10(model->lmath, score));
            for (j = 0; j <= i; ++j) {
                assert(wids[j] < model->n_counts[0]);
                fprintf(fh, "%s ", model->word_str[wids[j]]);
            }
            if (i < model->n-1)
                fprintf(fh, "%.4f", logmath_log_to_log10(model->lmath, bowt));
            fprintf(fh, "\n");
        }
    }
    fprintf(fh, "\n\\end\\\n");
    return fclose(fh);
}

static int
ngram_model_arpa_apply_weights(ngram_model_t *base, float32 lw,
                              float32 wip, float32 uw)
{
    ngram_model_arpa_t *model = (ngram_model_arpa_t *)base;
    lm3g_apply_weights(base, &model->lm3g, lw, wip, uw);
    return 0;
}

/* Lousy "templating" for things that are largely the same in DMP and
 * ARPA models, except for the bigram and trigram types and some
 * names. */
#define NGRAM_MODEL_TYPE ngram_model_arpa_t
#include "lm3g_templates.c"

static void
ngram_model_arpa_free(ngram_model_t *base)
{
    ngram_model_arpa_t *model = (ngram_model_arpa_t *)base;
    ckd_free(model->lm3g.unigrams);
    ckd_free(model->lm3g.bigrams);
    ckd_free(model->lm3g.trigrams);
    ckd_free(model->lm3g.prob2);
    ckd_free(model->lm3g.bo_wt2);
    ckd_free(model->lm3g.prob3);
    lm3g_tginfo_free(base, &model->lm3g);
    ckd_free(model->lm3g.tseg_base);
}

static ngram_funcs_t ngram_model_arpa_funcs = {
    ngram_model_arpa_free,          /* free */
    ngram_model_arpa_apply_weights, /* apply_weights */
    lm3g_template_score,            /* score */
    lm3g_template_raw_score,        /* raw_score */
    lm3g_template_add_ug,           /* add_ug */
    lm3g_template_flush,            /* flush */
    lm3g_template_iter,             /* iter */
    lm3g_template_mgrams,           /* mgrams */
    lm3g_template_successors,       /* successors */
    lm3g_template_iter_get,         /* iter_get */
    lm3g_template_iter_next,        /* iter_next */
    lm3g_template_iter_free         /* iter_free */
};