int32
ngram_model_init(ngram_model_t *base,
ngram_funcs_t *funcs,
logmath_t *lmath,
int32 n, int32 n_unigram)
{
base->refcount = 1;
base->funcs = funcs;
base->n = n;
/* If this was previously initialized... */
if (base->n_counts == NULL)
base->n_counts = ckd_calloc(3, sizeof(*base->n_counts));
/* Don't reset weights if logmath object hasn't changed. */
if (base->lmath != lmath) {
/* Set default values for weights. */
base->lw = 1.0;
base->log_wip = 0; /* i.e. 1.0 */
base->log_uw = 0; /* i.e. 1.0 */
base->log_uniform = logmath_log(lmath, 1.0 / n_unigram);
base->log_uniform_weight = logmath_get_zero(lmath);
base->log_zero = logmath_get_zero(lmath);
base->lmath = lmath;
}
/* Allocate or reallocate space for word strings. */
if (base->word_str) {
/* Free all previous word strings if they were allocated. */
if (base->writable) {
int32 i;
for (i = 0; i < base->n_words; ++i) {
ckd_free(base->word_str[i]);
base->word_str[i] = NULL;
}
}
base->word_str = ckd_realloc(base->word_str, n_unigram * sizeof(char *));
}
else
base->word_str = ckd_calloc(n_unigram, sizeof(char *));
/* NOTE: They are no longer case-insensitive since we are allowing
* other encodings for word strings. Beware. */
if (base->wid)
hash_table_empty(base->wid);
else
base->wid = hash_table_new(n_unigram, FALSE);
base->n_counts[0] = base->n_1g_alloc = base->n_words = n_unigram;
return 0;
}
void
acmod_grow_feat_buf(acmod_t *acmod, int nfr)
{
mfcc_t ***new_feat_buf;
new_feat_buf = feat_array_alloc(acmod->fcb, nfr);
if (acmod->n_feat_frame || acmod->grow_feat) {
memcpy(new_feat_buf[0][0], acmod->feat_buf[0][0],
(acmod->n_feat_alloc
* feat_dimension(acmod->fcb)
* sizeof(***acmod->feat_buf)));
}
feat_array_free(acmod->feat_buf);
acmod->framepos = ckd_realloc(acmod->framepos,
nfr * sizeof(*acmod->framepos));
acmod->feat_buf = new_feat_buf;
acmod->n_feat_alloc = nfr;
}
/**
* Add a word to the word string and ID mapping.
*/
int32
ngram_add_word_internal(ngram_model_t * model,
const char *word, int32 classid)
{
/* Check for hash collisions. */
int32 wid;
if (hash_table_lookup_int32(model->wid, word, &wid) == 0) {
E_WARN("Omit duplicate word '%s'\n", word);
return wid;
}
/* Take the next available word ID */
wid = model->n_words;
if (classid >= 0) {
wid = NGRAM_CLASSWID(wid, classid);
}
/* Reallocate word_str if necessary. */
if (model->n_words >= model->n_1g_alloc) {
model->n_1g_alloc += UG_ALLOC_STEP;
model->word_str = ckd_realloc(model->word_str,
sizeof(*model->word_str) *
model->n_1g_alloc);
}
/* Add the word string in the appropriate manner. */
/* Class words are always dynamically allocated. */
model->word_str[model->n_words] = ckd_salloc(word);
/* Now enter it into the hash table. */
if (hash_table_enter_int32
(model->wid, model->word_str[model->n_words], wid) != wid) {
E_ERROR
("Hash insertion failed for word %s => %p (should not happen)\n",
model->word_str[model->n_words], (void *) (long) (wid));
}
/* Increment number of words. */
++model->n_words;
return wid;
}
static size_t
strnappend(char **dest, size_t *dest_allocation,
const char *source, size_t n)
{
size_t source_len, required_allocation;
if (dest == NULL || dest_allocation == NULL)
return -1;
if (*dest == NULL && *dest_allocation != 0)
return -1;
if (source == NULL)
return *dest_allocation;
source_len = strlen(source);
if (n && n < source_len)
source_len = n;
required_allocation = (*dest ? strlen(*dest) : 0) + source_len + 1;
if (*dest_allocation < required_allocation) {
if (*dest_allocation == 0) {
*dest = ckd_calloc(required_allocation * 2, 1);
} else {
*dest = ckd_realloc(*dest, required_allocation * 2);
}
*dest_allocation = required_allocation * 2;
}
#import "OpenEarsStaticAnalysisToggle.h"
#ifdef STATICANALYZEDEPENDENCIES
#define __clang_analyzer__ 1
#endif
#if !defined(__clang_analyzer__) || defined(STATICANALYZEDEPENDENCIES)
#undef __clang_analyzer__
strncat(*dest, source, source_len);
#endif
return *dest_allocation;
}
int
fsg_model_word_add(fsg_model_t *fsg, char const *word)
{
int wid;
/* Search for an existing word matching this. */
wid = fsg_model_word_id(fsg, word);
/* If not found, add this to the vocab. */
if (wid == -1) {
wid = fsg->n_word;
if (fsg->n_word == fsg->n_word_alloc) {
fsg->n_word_alloc += 10;
fsg->vocab = ckd_realloc(fsg->vocab,
fsg->n_word_alloc * sizeof(*fsg->vocab));
if (fsg->silwords)
fsg->silwords = bitvec_realloc(fsg->silwords, fsg->n_word_alloc);
if (fsg->altwords)
fsg->altwords = bitvec_realloc(fsg->altwords, fsg->n_word_alloc);
}
++fsg->n_word;
fsg->vocab[wid] = ckd_salloc(word);
}
return wid;
}
lmset_t *
lmset_read_ctl(const char *ctlfile,
dict_t * dict,
float64 lw, float64 wip, float64 uw,
const char *lmdumpdir, logmath_t *logmath)
{
FILE *ctlfp;
FILE *tmp;
char lmfile[4096], lmname[4096], str[4096];
lmclass_set_t *lmclass_set;
lmclass_t **lmclass, *cl;
int32 n_lmclass, n_lmclass_used;
int32 i;
lm_t *lm;
lmset_t *lms = NULL;
tmp = NULL;
E_INFO("Reading LM control file '%s'\n", ctlfile);
if ((ctlfp = fopen(ctlfile, "r")) == NULL) {
E_ERROR_SYSTEM("Failed to open LM control file");
return NULL;
}
lmclass_set = lmclass_newset();
lms = (lmset_t *) ckd_calloc(1, sizeof(lmset_t));
lms->n_lm = 0;
lms->n_alloc_lm = 0;
if (fscanf(ctlfp, "%s", str) == 1) {
if (strcmp(str, "{") == 0) {
/* Load LMclass files */
while ((fscanf(ctlfp, "%s", str) == 1)
&& (strcmp(str, "}") != 0))
lmclass_set = lmclass_loadfile(lmclass_set, str, logmath);
if (strcmp(str, "}") != 0)
E_FATAL("Unexpected EOF(%s)\n", ctlfile);
if (fscanf(ctlfp, "%s", str) != 1)
str[0] = '\0';
}
}
else
str[0] = '\0';
/* Fill in dictionary word id information for each LMclass word */
for (cl = lmclass_firstclass(lmclass_set);
lmclass_isclass(cl); cl = lmclass_nextclass(lmclass_set, cl)) {
/*
For every words in the class, set the dictwid correctly
The following piece of code replace s2's kb_init_lmclass_dictwid (cl);
doesn't do any checking even the id is a bad dict id.
This only sets the information in the lmclass_set, but not
lm-2-dict or dict-2-lm map. In Sphinx 3, they are done in
wid_dict_lm_map in wid.c.
*/
lmclass_word_t *w;
int32 wid;
for (w = lmclass_firstword(cl); lmclass_isword(w);
w = lmclass_nextword(cl, w)) {
wid = dict_wordid(dict, lmclass_getword(w));
#if 0
E_INFO("In class %s, Word %s, wid %d\n", cl->name,
lmclass_getword(w), wid);
#endif
lmclass_set_dictwid(w, wid);
}
}
/* At this point if str[0] != '\0', we have an LM filename */
n_lmclass = lmclass_get_nclass(lmclass_set);
lmclass = (lmclass_t **) ckd_calloc(n_lmclass, sizeof(lmclass_t *));
E_INFO("Number of LM class specified %d in file %s\n", n_lmclass,
ctlfile);
/* Read in one LM at a time */
while (str[0] != '\0') {
strcpy(lmfile, str);
if (fscanf(ctlfp, "%s", lmname) != 1)
E_FATAL("LMname missing after LMFileName '%s'\n", lmfile);
n_lmclass_used = 0;
if (fscanf(ctlfp, "%s", str) == 1) {
if (strcmp(str, "{") == 0) {
while ((fscanf(ctlfp, "%s", str) == 1) &&
(strcmp(str, "}") != 0)) {
if (n_lmclass_used >= n_lmclass) {
E_FATAL("Too many LM classes specified for '%s'\n",
lmfile);
}
lmclass[n_lmclass_used] =
lmclass_get_lmclass(lmclass_set, str);
if (!(lmclass_isclass(lmclass[n_lmclass_used])))
E_FATAL("LM class '%s' not found\n", str);
n_lmclass_used++;
}
if (strcmp(str, "}") != 0)
E_FATAL("Unexpected EOF(%s)\n", ctlfile);
if (fscanf(ctlfp, "%s", str) != 1)
str[0] = '\0';
}
}
else
str[0] = '\0';
lm = (lm_t *) lm_read_advance(lmfile, lmname, lw, wip, uw,
dict_size(dict), NULL, 1, logmath, FALSE, FALSE);
if (n_lmclass_used > 0) {
E_INFO("Did I enter here?\n");
lm_build_lmclass_info(lm, lw, uw, wip, n_lmclass_used,
lmclass);
}
if (lms->n_lm == lms->n_alloc_lm) {
lms->lmarray =
(lm_t **) ckd_realloc(lms->lmarray,
(lms->n_alloc_lm +
LM_ALLOC_BLOCK) * sizeof(lm_t *));
lms->n_alloc_lm += LM_ALLOC_BLOCK;
}
lms->lmarray[lms->n_lm] = lm;
lms->n_lm += 1;
E_INFO("%d %d\n", lms->n_alloc_lm, lms->n_lm);
}
assert(lms);
assert(lms->lmarray);
E_INFO("No. of LM set allocated %d, no. of LM %d \n", lms->n_alloc_lm,
lms->n_lm);
if (dict != NULL) {
for (i = 0; i < lms->n_lm; i++) {
assert(lms->lmarray[i]);
assert(dict);
if ((lms->lmarray[i]->dict2lmwid =
wid_dict_lm_map(dict, lms->lmarray[i], lw)) == NULL)
E_FATAL
("Dict/LM word-id mapping failed for LM index %d, named %s\n",
i, lmset_idx_to_name(lms, i));
}
}
else {
E_FATAL
("Dict is specified to be NULL (dict_init is not called before lmset_read_lm?), dict2lmwid is not built inside lmset_read_lm\n");
}
ckd_free(lmclass_set);
ckd_free(lmclass);
fclose(ctlfp);
return lms;
}
int32
feat_s2mfc2feat(feat_t * fcb, const char *file, const char *dir, const char *cepext,
int32 sf, int32 ef, mfcc_t *** feat, int32 maxfr)
{
char *path;
char *ps = "/";
int32 win, nfr;
int32 file_length, cepext_length, path_length = 0;
mfcc_t **mfc;
if (fcb->cepsize <= 0) {
E_ERROR("Bad cepsize: %d\n", fcb->cepsize);
return -1;
}
if (cepext == 0)
cepext = "";
/*
* Create mfc filename, combining file, dir and extension if
* necessary
*/
/*
* First we decide about the path. If dir is defined, then use
* it. Otherwise assume the filename already contains the path.
*/
if (dir == 0) {
dir = "";
ps = "";
/*
* This is not true but some 3rd party apps
* may parse the output explicitly checking for this line
*/
E_INFO("At directory . (current directory)\n");
}
else {
E_INFO("At directory %s\n", dir);
/*
* Do not forget the path separator!
*/
path_length += strlen(dir) + 1;
}
/*
* Include cepext, if it's not already part of the filename.
*/
file_length = strlen(file);
cepext_length = strlen(cepext);
if ((file_length > cepext_length)
&& (strcmp(file + file_length - cepext_length, cepext) == 0)) {
cepext = "";
cepext_length = 0;
}
/*
* Do not forget the '\0'
*/
path_length += file_length + cepext_length + 1;
path = (char*) ckd_calloc(path_length, sizeof(char));
#ifdef HAVE_SNPRINTF
/*
* Paranoia is our best friend...
*/
while ((file_length = snprintf(path, path_length, "%s%s%s%s", dir, ps, file, cepext)) > path_length) {
path_length = file_length;
path = (char*) ckd_realloc(path, path_length * sizeof(char));
}
#else
#ifndef POCKETSPHINX_NET
sprintf(path, "%s%s%s%s", dir, ps, file, cepext);
#else
strcpy(path,dir);
strcat(path,ps);
strcat(path,file);
strcat(path,cepext);
#endif
#endif
win = feat_window_size(fcb);
/* Pad maxfr with win, so we read enough raw feature data to
* calculate the requisite number of dynamic features. */
if (maxfr >= 0)
maxfr += win * 2;
if (feat != 0) {
/* Read mfc file including window or padding if necessary. */
nfr = feat_s2mfc_read_norm_pad(fcb, path, win, sf, ef, &mfc, maxfr, fcb->cepsize);
ckd_free(path);
if (nfr < 0) {
ckd_free_2d((void **) mfc);
return -1;
}
/* Actually compute the features */
feat_compute_utt(fcb, mfc, nfr, win, feat);
ckd_free_2d((void **) mfc);
}
else {
/* Just calculate the number of frames we would need. */
nfr = feat_s2mfc_read_norm_pad(fcb, path, win, sf, ef, 0, maxfr, fcb->cepsize);
ckd_free(path);
if (nfr < 0)
return nfr;
}
return (nfr - win * 2);
}
cmd_ln_t *
cmd_ln_parse_file_r(cmd_ln_t *inout_cmdln, const arg_t * defn, const char *filename, int32 strict)
{
FILE *file;
int argc;
int argv_size;
char *str;
int arg_max_length = 512;
int len = 0;
int quoting, ch;
char **f_argv;
int rv = 0;
const char separator[] = " \t\r\n";
if ((file = fopen(filename, "r")) == NULL) {
E_ERROR("Cannot open configuration file %s for reading\n",
filename);
return NULL;
}
ch = fgetc(file);
/* Skip to the next interesting character */
for (; ch != EOF && strchr(separator, ch); ch = fgetc(file)) ;
if (ch == EOF) {
fclose(file);
return NULL;
}
/*
* Initialize default argv, argc, and argv_size.
*/
argv_size = 10;
argc = 0;
f_argv = ckd_calloc(argv_size, sizeof(char *));
/* Silently make room for \0 */
str = ckd_calloc(arg_max_length + 1, sizeof(char));
quoting = 0;
do {
/* Handle arguments that are commented out */
if (len == 0 && argc % 2 == 0) {
while (ch == '#') {
/* Skip everything until newline */
for (ch = fgetc(file); ch != EOF && ch != '\n'; ch = fgetc(file)) ;
/* Skip to the next interesting character */
for (ch = fgetc(file); ch != EOF && strchr(separator, ch); ch = fgetc(file)) ;
}
/* Check if we are at the last line (without anything interesting in it) */
if (ch == EOF)
break;
}
/* Handle quoted arguments */
if (ch == '"' || ch == '\'') {
if (quoting == ch) /* End a quoted section with the same type */
quoting = 0;
else if (quoting) {
E_ERROR("Nesting quotations is not supported!\n");
rv = 1;
break;
}
else
quoting = ch; /* Start a quoted section */
}
else if (ch == EOF || (!quoting && strchr(separator, ch))) {
/* Reallocate argv so it is big enough to contain all the arguments */
if (argc >= argv_size) {
char **tmp_argv;
if (!(tmp_argv =
ckd_realloc(f_argv, argv_size * 2 * sizeof(char *)))) {
rv = 1;
break;
}
f_argv = tmp_argv;
argv_size *= 2;
}
/* Add the string to the list of arguments */
f_argv[argc] = ckd_salloc(str);
len = 0;
str[0] = '\0';
argc++;
if (quoting)
E_WARN("Unclosed quotation, having EOF close it...\n");
/* Skip to the next interesting character */
for (; ch != EOF && strchr(separator, ch); ch = fgetc(file)) ;
if (ch == EOF)
break;
/* We already have the next character */
continue;
}
else {
if (len >= arg_max_length) {
#import "OpenEarsStaticAnalysisToggle.h"
#ifdef STATICANALYZEDEPENDENCIES
#define __clang_analyzer__ 1
#endif
#if !defined(__clang_analyzer__) || defined(STATICANALYZEDEPENDENCIES)
#undef __clang_analyzer__
/* Make room for more chars (including the \0 !) */
char *tmp_str = str;
if ((tmp_str = ckd_realloc(str, (1 + arg_max_length * 2) * sizeof(char))) == NULL) {
rv = 1;
break;
}
str = tmp_str;
arg_max_length *= 2;
#endif
}
/* Add the char to the argument string */
str[len++] = ch;
/* Always null terminate */
str[len] = '\0';
}
ch = fgetc(file);
} while (1);
fclose(file);
ckd_free(str);
if (rv) {
for (ch = 0; ch < argc; ++ch)
ckd_free(f_argv[ch]);
ckd_free(f_argv);
return NULL;
}
return parse_options(inout_cmdln, defn, argc, f_argv, strict);
}
int
ps_decoder_test(cmd_ln_t *config, char const *sname, char const *expected)
{
ps_decoder_t *ps;
mfcc_t **cepbuf;
FILE *rawfh;
int16 *buf;
int16 const *bptr;
size_t nread;
size_t nsamps;
int32 nfr, i, score, prob;
char const *hyp;
char const *uttid;
double n_speech, n_cpu, n_wall;
ps_seg_t *seg;
TEST_ASSERT(ps = ps_init(config));
/* Test it first with pocketsphinx_decode_raw() */
TEST_ASSERT(rawfh = fopen(DATADIR "/goforward.raw", "rb"));
ps_decode_raw(ps, rawfh, "goforward", -1);
hyp = ps_get_hyp(ps, &score, &uttid);
prob = ps_get_prob(ps, &uttid);
printf("%s (%s): %s (%d, %d)\n", sname, uttid, hyp, score, prob);
TEST_EQUAL(0, strcmp(hyp, expected));
TEST_ASSERT(prob <= 0);
ps_get_utt_time(ps, &n_speech, &n_cpu, &n_wall);
printf("%.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
printf("%.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
/* Test it with ps_process_raw() */
clearerr(rawfh);
fseek(rawfh, 0, SEEK_END);
nsamps = ftell(rawfh) / sizeof(*buf);
fseek(rawfh, 0, SEEK_SET);
TEST_EQUAL(0, ps_start_utt(ps, NULL));
nsamps = 2048;
buf = ckd_calloc(nsamps, sizeof(*buf));
while (!feof(rawfh)) {
nread = fread(buf, sizeof(*buf), nsamps, rawfh);
ps_process_raw(ps, buf, nread, FALSE, FALSE);
}
TEST_EQUAL(0, ps_end_utt(ps));
hyp = ps_get_hyp(ps, &score, &uttid);
prob = ps_get_prob(ps, &uttid);
printf("%s (%s): %s (%d, %d)\n", sname, uttid, hyp, score, prob);
TEST_EQUAL(0, strcmp(uttid, "000000000"));
TEST_EQUAL(0, strcmp(hyp, expected));
ps_get_utt_time(ps, &n_speech, &n_cpu, &n_wall);
printf("%.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
printf("%.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
/* Now read the whole file and produce an MFCC buffer. */
clearerr(rawfh);
fseek(rawfh, 0, SEEK_END);
nsamps = ftell(rawfh) / sizeof(*buf);
fseek(rawfh, 0, SEEK_SET);
bptr = buf = ckd_realloc(buf, nsamps * sizeof(*buf));
TEST_EQUAL(nsamps, fread(buf, sizeof(*buf), nsamps, rawfh));
fe_process_frames(ps->acmod->fe, &bptr, &nsamps, NULL, &nfr);
cepbuf = ckd_calloc_2d(nfr + 1,
fe_get_output_size(ps->acmod->fe),
sizeof(**cepbuf));
fe_start_utt(ps->acmod->fe);
fe_process_frames(ps->acmod->fe, &bptr, &nsamps, cepbuf, &nfr);
fe_end_utt(ps->acmod->fe, cepbuf[nfr], &i);
/* Decode it with process_cep() */
TEST_EQUAL(0, ps_start_utt(ps, NULL));
for (i = 0; i < nfr; ++i) {
ps_process_cep(ps, cepbuf + i, 1, FALSE, FALSE);
}
TEST_EQUAL(0, ps_end_utt(ps));
hyp = ps_get_hyp(ps, &score, &uttid);
prob = ps_get_prob(ps, &uttid);
printf("%s (%s): %s (%d, %d)\n", sname, uttid, hyp, score, prob);
TEST_EQUAL(0, strcmp(uttid, "000000001"));
TEST_EQUAL(0, strcmp(hyp, expected));
TEST_ASSERT(prob <= 0);
for (seg = ps_seg_iter(ps, &score); seg;
seg = ps_seg_next(seg)) {
char const *word;
int sf, ef;
int32 post, lscr, ascr, lback;
word = ps_seg_word(seg);
ps_seg_frames(seg, &sf, &ef);
post = ps_seg_prob(seg, &ascr, &lscr, &lback);
printf("%s (%d:%d) P(w|o) = %f ascr = %d lscr = %d lback = %d\n", word, sf, ef,
logmath_exp(ps_get_logmath(ps), post), ascr, lscr, lback);
TEST_ASSERT(post <= 2); // Due to numerical errors with float it sometimes could go out of 0
}
ps_get_utt_time(ps, &n_speech, &n_cpu, &n_wall);
printf("%.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
printf("%.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
ps_get_all_time(ps, &n_speech, &n_cpu, &n_wall);
printf("TOTAL: %.2f seconds speech, %.2f seconds CPU, %.2f seconds wall\n",
n_speech, n_cpu, n_wall);
printf("TOTAL: %.2f xRT (CPU), %.2f xRT (elapsed)\n",
n_cpu / n_speech, n_wall / n_speech);
fclose(rawfh);
ps_free(ps);
cmd_ln_free_r(config);
ckd_free_2d(cepbuf);
ckd_free(buf);
return 0;
}
int
main(int argc, char *argv[])
{
acmod_t *acmod;
logmath_t *lmath;
cmd_ln_t *config;
FILE *rawfh;
int16 *buf;
int16 const *bptr;
mfcc_t **cepbuf, **cptr;
size_t nread, nsamps;
int nfr;
int frame_counter;
int bestsen1[270];
lmath = logmath_init(1.0001, 0, 0);
config = cmd_ln_init(NULL, ps_args(), TRUE,
"-featparams", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/feat.params",
"-mdef", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/mdef",
"-mean", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/means",
"-var", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/variances",
"-tmat", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/transition_matrices",
"-sendump", MODELDIR "/hmm/en_US/hub4wsj_sc_8k/sendump",
"-compallsen", "true",
"-cmn", "prior",
"-tmatfloor", "0.0001",
"-mixwfloor", "0.001",
"-varfloor", "0.0001",
"-mmap", "no",
"-topn", "4",
"-ds", "1",
"-input_endian", "little",
"-samprate", "16000", NULL);
TEST_ASSERT(config);
TEST_ASSERT(acmod = acmod_init(config, lmath, NULL, NULL));
cmn_prior_set(acmod->fcb->cmn_struct, prior);
nsamps = 2048;
frame_counter = 0;
buf = ckd_calloc(nsamps, sizeof(*buf));
TEST_ASSERT(rawfh = fopen(DATADIR "/goforward.raw", "rb"));
TEST_EQUAL(0, acmod_start_utt(acmod));
E_INFO("Incremental(2048):\n");
while (!feof(rawfh)) {
nread = fread(buf, sizeof(*buf), nsamps, rawfh);
bptr = buf;
while ((nfr = acmod_process_raw(acmod, &bptr, &nread, FALSE)) > 0 || nread > 0) {
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
bestsen1[frame_counter] = best_score;
++frame_counter;
frame_idx = -1;
}
}
}
TEST_EQUAL(0, acmod_end_utt(acmod));
nread = 0;
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
bestsen1[frame_counter] = best_score;
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Now try to process the whole thing at once. */
E_INFO("Whole utterance:\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
nsamps = ftell(rawfh) / sizeof(*buf);
clearerr(rawfh);
fseek(rawfh, 0, SEEK_SET);
buf = ckd_realloc(buf, nsamps * sizeof(*buf));
TEST_EQUAL(nsamps, fread(buf, sizeof(*buf), nsamps, rawfh));
bptr = buf;
TEST_EQUAL(0, acmod_start_utt(acmod));
acmod_process_raw(acmod, &bptr, &nsamps, TRUE);
TEST_EQUAL(0, acmod_end_utt(acmod));
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Now process MFCCs and make sure we get the same results. */
cepbuf = ckd_calloc_2d(frame_counter,
fe_get_output_size(acmod->fe),
sizeof(**cepbuf));
fe_start_utt(acmod->fe);
nsamps = ftell(rawfh) / sizeof(*buf);
bptr = buf;
nfr = frame_counter;
fe_process_frames(acmod->fe, &bptr, &nsamps, cepbuf, &nfr);
fe_end_utt(acmod->fe, cepbuf[frame_counter-1], &nfr);
E_INFO("Incremental(MFCC):\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
TEST_EQUAL(0, acmod_start_utt(acmod));
cptr = cepbuf;
nfr = frame_counter;
frame_counter = 0;
while ((acmod_process_cep(acmod, &cptr, &nfr, FALSE)) > 0) {
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
++frame_counter;
frame_idx = -1;
}
}
TEST_EQUAL(0, acmod_end_utt(acmod));
nfr = 0;
acmod_process_cep(acmod, &cptr, &nfr, FALSE);
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
TEST_EQUAL(frame_counter, frame_idx);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
++frame_counter;
frame_idx = -1;
}
}
/* Note that we have to process the whole thing again because
* !#@$@ s2mfc2feat modifies its argument (not for long) */
fe_start_utt(acmod->fe);
nsamps = ftell(rawfh) / sizeof(*buf);
bptr = buf;
nfr = frame_counter;
fe_process_frames(acmod->fe, &bptr, &nsamps, cepbuf, &nfr);
fe_end_utt(acmod->fe, cepbuf[frame_counter-1], &nfr);
E_INFO("Whole utterance (MFCC):\n");
cmn_prior_set(acmod->fcb->cmn_struct, prior);
TEST_EQUAL(0, acmod_start_utt(acmod));
cptr = cepbuf;
nfr = frame_counter;
acmod_process_cep(acmod, &cptr, &nfr, TRUE);
TEST_EQUAL(0, acmod_end_utt(acmod));
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
E_INFO("Rewound (MFCC):\n");
TEST_EQUAL(0, acmod_rewind(acmod));
{
int16 const *senscr;
int16 best_score;
int frame_idx = -1, best_senid;
frame_counter = 0;
while (acmod->n_feat_frame > 0) {
senscr = acmod_score(acmod, &frame_idx);
acmod_advance(acmod);
best_score = acmod_best_score(acmod, &best_senid);
E_INFO("Frame %d best senone %d score %d\n",
frame_idx, best_senid, best_score);
if (frame_counter < 190)
TEST_EQUAL_LOG(best_score, bestsen1[frame_counter]);
TEST_EQUAL(frame_counter, frame_idx);
++frame_counter;
frame_idx = -1;
}
}
/* Clean up, go home. */
ckd_free_2d(cepbuf);
fclose(rawfh);
ckd_free(buf);
acmod_free(acmod);
logmath_free(lmath);
cmd_ln_free_r(config);
return 0;
}
s3wid_t
dict_add_word(dict_t * d, char const *word, s3cipid_t const * p, int32 np)
{
int32 len;
dictword_t *wordp;
s3wid_t newwid;
char *wword;
if (d->n_word >= d->max_words) {
E_INFO("Reallocating to %d KiB for word entries\n",
(d->max_words + S3DICT_INC_SZ) * sizeof(dictword_t) / 1024);
d->word =
(dictword_t *) ckd_realloc(d->word,
(d->max_words +
S3DICT_INC_SZ) * sizeof(dictword_t));
d->max_words = d->max_words + S3DICT_INC_SZ;
}
wordp = d->word + d->n_word;
wordp->word = (char *) ckd_salloc(word); /* Freed in dict_free */
/* Associate word string with d->n_word in hash table */
if (hash_table_enter_int32(d->ht, wordp->word, d->n_word) != d->n_word) {
ckd_free(wordp->word);
wordp->word = NULL;
return BAD_S3WID;
}
/* Fill in word entry, and set defaults */
if (p && (np > 0)) {
wordp->ciphone = (s3cipid_t *) ckd_malloc(np * sizeof(s3cipid_t)); /* Freed in dict_free */
memcpy(wordp->ciphone, p, np * sizeof(s3cipid_t));
wordp->pronlen = np;
}
else {
wordp->ciphone = NULL;
wordp->pronlen = 0;
}
wordp->alt = BAD_S3WID;
wordp->basewid = d->n_word;
/* Determine base/alt wids */
wword = ckd_salloc(word);
if ((len = dict_word2basestr(wword)) > 0) {
int32 w;
/* Truncated to a baseword string; find its ID */
if (hash_table_lookup_int32(d->ht, wword, &w) < 0) {
E_ERROR("Missing base word for: %s\n", word);
ckd_free(wword);
ckd_free(wordp->word);
wordp->word = NULL;
return BAD_S3WID;
}
/* Link into alt list */
wordp->basewid = w;
wordp->alt = d->word[w].alt;
d->word[w].alt = d->n_word;
}
ckd_free(wword);
newwid = d->n_word++;
return newwid;
}
static int32
dict_read(FILE * fp, dict_t * d)
{
lineiter_t *li;
char **wptr;
s3cipid_t *p;
int32 lineno, nwd;
s3wid_t w;
int32 i, maxwd;
size_t stralloc, phnalloc;
maxwd = 512;
p = (s3cipid_t *) ckd_calloc(maxwd + 4, sizeof(*p));
wptr = (char **) ckd_calloc(maxwd, sizeof(char *)); /* Freed below */
lineno = 0;
stralloc = phnalloc = 0;
for (li = lineiter_start(fp); li; li = lineiter_next(li)) {
lineno++;
if (0 == strncmp(li->buf, "##", 2)
|| 0 == strncmp(li->buf, ";;", 2))
continue;
if ((nwd = str2words(li->buf, wptr, maxwd)) < 0) {
/* Increase size of p, wptr. */
nwd = str2words(li->buf, NULL, 0);
assert(nwd > maxwd); /* why else would it fail? */
maxwd = nwd;
p = (s3cipid_t *) ckd_realloc(p, (maxwd + 4) * sizeof(*p));
wptr = (char **) ckd_realloc(wptr, maxwd * sizeof(*wptr));
}
if (nwd == 0) /* Empty line */
continue;
/* wptr[0] is the word-string and wptr[1..nwd-1] the pronunciation sequence */
if (nwd == 1) {
E_ERROR("Line %d: No pronunciation for word %s; ignored\n",
lineno, wptr[0]);
continue;
}
/* Convert pronunciation string to CI-phone-ids */
for (i = 1; i < nwd; i++) {
p[i - 1] = dict_ciphone_id(d, wptr[i]);
if (NOT_S3CIPID(p[i - 1])) {
E_ERROR("Line %d: Bad ciphone: %s; word %s ignored\n",
lineno, wptr[i], wptr[0]);
break;
}
}
if (i == nwd) { /* All CI-phones successfully converted to IDs */
w = dict_add_word(d, wptr[0], p, nwd - 1);
if (NOT_S3WID(w))
E_ERROR
("Line %d: dict_add_word (%s) failed (duplicate?); ignored\n",
lineno, wptr[0]);
else {
stralloc += strlen(d->word[w].word);
phnalloc += d->word[w].pronlen * sizeof(s3cipid_t);
}
}
}
E_INFO("Allocated %d KiB for strings, %d KiB for phones\n",
(int)stralloc / 1024, (int)phnalloc / 1024);
ckd_free(p);
ckd_free(wptr);
return 0;
}
void kb_setlm(char* lmname,kb_t* kb)
{
lmset_t* lms;
kbcore_t* kbc=NULL;
int i;
int j;
int n;
/* s3wid_t dictid;*/
kbc=kb->kbcore;
lms=kbc->lmset;
E_INFO("Inside kb_setlm\n");
kbc->lm=NULL;
for(j=0;j<kb->n_lextree;j++){
kb->ugtree[j]=NULL;
}
E_INFO("Inside kb_setlm\n");
if(lms!=NULL || cmd_ln_str("-lmctlfn")){
for(i=0;i<kbc->n_lm;i++){
if(!strcmp(lmname,lms[i].name)){
/* Point the current lm to a particular lm */
kbc->lm=lms[i].lm;
for(j=0;j<kb->n_lextree;j++){
kb->ugtree[j]=kb->ugtreeMulti[i*kb->n_lextree+j];
}
break;
}
}
if(kbc->lm==NULL){
E_ERROR("LM name %s cannot be found! Fall back to use base LM model\n",lmname);
kbc->lm=lms[0].lm;
for(j=0;j<kb->n_lextree;j++){
kb->ugtree[j]=kb->ugtreeMulti[j];
}
}
}
E_INFO("Current LM name %s.\n",lms[i].name);
/* if((kb->vithist->lms2vh_root=
(vh_lms2vh_t**)ckd_realloc(kb->vithist->lms2vh_root,
lm_n_ug(kbc->lm)*sizeof(vh_lms2vh_t *)
))==NULL)
{
E_FATAL("failed to allocate memory for vithist\n");
}*/
n = ((kb->ugtree[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
n /= kb->hmm_hist_binsize;
kb->hmm_hist_bins = n+1;
kb->hmm_hist = (int32 *) ckd_realloc (kb->hmm_hist,(n+1)*sizeof(int32)); /* Really no need for +1 */
E_INFO("Current LM name %s\n",lms[i].name);
E_INFO("LM ug size %d\n",kbc->lm->n_ug);
E_INFO("LM bg size %d\n",kbc->lm->n_bg);
E_INFO("LM tg size %d\n",kbc->lm->n_tg);
E_INFO("HMM history bin size %d\n", n+1);
for(j=0;j<kb->n_lextree;j++){
E_INFO("Lextrees(%d), %d nodes(ug)\n",
kb->n_lextree, lextree_n_node(kb->ugtree[j]));
}
/* for (n = 0; n < dict_size(kbcore_dict(kbc)); n++){
E_INFO("Index %d, map %d word %s\n",n,kbc->lm->dict2lmwid[n],dict_wordstr(kbcore_dict(kbc),n));
}*/
}
int
ngram_model_recode(ngram_model_t *model, const char *from, const char *to)
{
iconv_t ic;
char *outbuf;
size_t maxlen;
int i, writable;
hash_table_t *new_wid;
/* FIXME: Need to do a special case thing for the GB-HEX encoding
* used in Sphinx3 Mandarin models. */
if ((ic = iconv_open(to, from)) == (iconv_t)-1) {
E_ERROR_SYSTEM("iconv_open() failed");
return -1;
}
/* iconv(3) is a piece of crap and won't accept a NULL out buffer,
* unlike wcstombs(3). So we have to either call it over and over
* again until our buffer is big enough, or call it with a huge
* buffer and then copy things back to the output. We will use a
* mix of these two approaches here. We'll keep a single big
* buffer around, and expand it as necessary.
*/
maxlen = 0;
for (i = 0; i < model->n_words; ++i) {
if (strlen(model->word_str[i]) > maxlen)
maxlen = strlen(model->word_str[i]);
}
/* Were word strings already allocated? */
writable = model->writable;
/* Either way, we are going to allocate some word strings. */
model->writable = TRUE;
/* Really should be big enough except for pathological cases. */
maxlen = maxlen * sizeof(int) + 15;
outbuf = ckd_calloc(maxlen, 1);
/* And, don't forget, we need to rebuild the word to unigram ID
* mapping. */
new_wid = hash_table_new(model->n_words, FALSE);
for (i = 0; i < model->n_words; ++i) {
ICONV_CONST char *in;
char *out;
size_t inleft, outleft, result;
start_conversion:
in = (ICONV_CONST char *)model->word_str[i];
/* Yes, this assumes that we don't have any NUL bytes. */
inleft = strlen(in);
out = outbuf;
outleft = maxlen;
while ((result = iconv(ic, &in, &inleft, &out, &outleft)) == (size_t)-1) {
if (errno != E2BIG) {
/* FIXME: if we already converted any words, then they
* are going to be in an inconsistent state. */
E_ERROR_SYSTEM("iconv() failed");
ckd_free(outbuf);
hash_table_free(new_wid);
return -1;
}
/* Reset the internal state of conversion. */
iconv(ic, NULL, NULL, NULL, NULL);
/* Make everything bigger. */
maxlen *= 2;
out = outbuf = ckd_realloc(outbuf, maxlen);
/* Reset the input pointers. */
in = (ICONV_CONST char *)model->word_str[i];
inleft = strlen(in);
}
/* Now flush a shift-out sequence, if any. */
if ((result = iconv(ic, NULL, NULL, &out, &outleft)) == (size_t)-1) {
if (errno != E2BIG) {
/* FIXME: if we already converted any words, then they
* are going to be in an inconsistent state. */
E_ERROR_SYSTEM("iconv() failed (state reset sequence)");
ckd_free(outbuf);
hash_table_free(new_wid);
return -1;
}
/* Reset the internal state of conversion. */
iconv(ic, NULL, NULL, NULL, NULL);
/* Make everything bigger. */
maxlen *= 2;
outbuf = ckd_realloc(outbuf, maxlen);
/* Be very evil. */
goto start_conversion;
}
result = maxlen - outleft;
/* Okay, that was hard, now let's go shopping. */
if (writable) {
/* Grow or shrink the output string as necessary. */
model->word_str[i] = ckd_realloc(model->word_str[i], result + 1);
model->word_str[i][result] = '\0';
}
else {
/* It actually was not allocated previously, so do that now. */
model->word_str[i] = ckd_calloc(result + 1, 1);
}
/* Copy the new thing in. */
memcpy(model->word_str[i], outbuf, result);
/* Now update the hash table. We might have terrible
* collisions if a non-reversible conversion was requested.,
* so warn about them. */
if (hash_table_enter_int32(new_wid, model->word_str[i], i) != i) {
E_WARN("Duplicate word in dictionary after conversion: %s\n",
model->word_str[i]);
}
}
ckd_free(outbuf);
iconv_close(ic);
/* Swap out the hash table. */
hash_table_free(model->wid);
model->wid = new_wid;
return 0;
}
s3wid_t
dict_add_word(dict_t * d, char *word, s3cipid_t * p, int32 np)
{
int32 len;
dictword_t *wordp;
s3wid_t newwid;
if (d->n_word >= d->max_words) {
E_INFO
("Dictionary max size (%d) exceeded; reallocate another entries %d \n",
d->max_words, DICT_INC_SZ);
d->word =
(dictword_t *) ckd_realloc(d->word,
(d->max_words +
DICT_INC_SZ) * sizeof(dictword_t));
d->max_words = d->max_words + DICT_INC_SZ;
return (BAD_S3WID);
}
wordp = d->word + d->n_word;
wordp->word = (char *) ckd_salloc(word); /* Freed in dict_free */
/* Associate word string with d->n_word in hash table */
if (hash_table_enter(d->ht, wordp->word, (void *)(long)d->n_word) != (void *)(long)d->n_word) {
ckd_free(wordp->word);
return (BAD_S3WID);
}
/* Fill in word entry, and set defaults */
if (p && (np > 0)) {
wordp->ciphone = (s3cipid_t *) ckd_malloc(np * sizeof(s3cipid_t)); /* Freed in dict_free */
memcpy(wordp->ciphone, p, np * sizeof(s3cipid_t));
wordp->pronlen = np;
}
else {
wordp->ciphone = NULL;
wordp->pronlen = 0;
}
wordp->alt = BAD_S3WID;
wordp->basewid = d->n_word;
wordp->n_comp = 0;
wordp->comp = NULL;
/* Determine base/alt wids */
if ((len = dict_word2basestr(word)) > 0) {
void *val;
s3wid_t w;
/* Truncated to a baseword string; find its ID */
if (hash_table_lookup(d->ht, word, &val) < 0) {
word[len] = '('; /* Get back the original word */
E_FATAL("Missing base word for: %s\n", word);
}
else
word[len] = '('; /* Get back the original word */
/* Link into alt list */
w = (s3wid_t)(long)val;
wordp->basewid = w;
wordp->alt = d->word[w].alt;
d->word[w].alt = d->n_word;
}
newwid = d->n_word++;
return (newwid);
}