/**
* <JA>
* 音声波形データから MFCC パラメータを抽出する．
*
* @param speech [in] 音声波形データ
* @param speechlen [in] @a speech の長さ（単位：サンプル数）
*
* @return 新たに割り付けられ抽出パラメータベクトルが格納されている
* パラメータ構造体へのポインタを返す．
* </JA>
* <EN>
* Extract MFCC parameters with sentence CMN from given waveform.
*
* @param speech [in] buffer of speech waveform
* @param speechlen [in] length of @a speech in samples
*
* @return pointer to newly allocated parameter structure data with extracted
* MFCC vector sequence.
* </EN>
*/
HTK_Param *new_wav2mfcc(SP16 speech[], int speechlen)
{
HTK_Param *param;
int framenum;
int i;
int len;
if (ssload_filename && ssbuf == NULL) {
/* load noise spectrum for spectral subtraction from file (once) */
if ((ssbuf = new_SS_load_from_file(ssload_filename, &sslen)) == NULL) {
j_error("Error: failed to read \"%s\"\n", ssload_filename);
}
}
if (sscalc) {
/* compute noise spectrum from head silence for each input */
len = sscalc_len * para.smp_freq / 1000;
if (len > speechlen) len = speechlen;
#ifdef SSDEBUG
printf("[%d]\n", len);
#endif
ssbuf = new_SS_calculate(speech, len, para, &sslen);
}
#ifdef SSDEBUG
{
int i;
for(i=0;i<sslen;i++) {
printf("%d: %f\n", i, ssbuf[i]);
}
}
#endif
/* calculate frame length from speech length, frame size and frame shift */
framenum = (int)((speechlen - para.framesize) / para.frameshift) + 1;
if (framenum < 1) {
j_printerr("input too short (%d samples), ignored\n", speechlen);
return NULL;
}
/* malloc new param */
param = new_param();
param->parvec = (VECT **)mymalloc(sizeof(VECT *) * framenum);
for(i=0;i<framenum;i++) {
param->parvec[i] = (VECT *)mymalloc(sizeof(VECT) * para.veclen);
}
/* make MFCC from speech data */
Wav2MFCC(speech, param->parvec, para, speechlen, ssbuf, sslen);
/* set miscellaneous parameters */
param->header.samplenum = framenum;
param->header.wshift = para.smp_period * para.frameshift;
param->header.sampsize = para.veclen * sizeof(VECT); /* not compressed */
param->header.samptype = F_MFCC;
if (para.delta) param->header.samptype |= F_DELTA;
if (para.acc) param->header.samptype |= F_ACCL;
if (para.energy) param->header.samptype |= F_ENERGY;
if (para.c0) param->header.samptype |= F_ZEROTH;
if (para.absesup) param->header.samptype |= F_ENERGY_SUP;
if (para.cmn) param->header.samptype |= F_CEPNORM;
param->veclen = para.veclen;
param->samplenum = framenum;
return param;
}
int
main(int argc, char *argv[])
{
Recog *recog;
Jconf *jconf;
float *ss;
MFCCWork *wrk;
/* create instance */
recog = j_recog_new();
jconf = j_jconf_new();
recog->jconf = jconf;
/* set application-specific additional options */
j_add_option("-freq", 1, 1, "sampling freq in Hz", opt_freq);
j_add_option("-len", 1, 1, "record length in msec", opt_len);
j_add_option("-h", 0, 0, "display this help", opt_help);
j_add_option("-help", 0, 0, "display this help", opt_help);
j_add_option("--help", 0, 0, "display this help", opt_help);
/* when no argument, output help and exit */
if (argc <= 1) {
opt_help(jconf, NULL, 0);
return 0;
}
/* regard last arg as filename */
if (strmatch(argv[argc-1], "-")) {
stout = TRUE;
} else {
filename = argv[argc-1];
}
/* set default as same as "-input mic" */
jconf->input.type = INPUT_WAVEFORM;
jconf->input.speech_input = SP_MIC;
jconf->input.device = SP_INPUT_DEFAULT;
/* process config and load them */
if (j_config_load_args(jconf, argc-1, argv) == -1) {
fprintf(stderr, "Error reading arguments\n");
return -1;
}
/* force some default values */
jconf->detect.silence_cut = 0; /* disable silence cut */
jconf->preprocess.strip_zero_sample = TRUE; /* strip zero samples */
jconf->detect.level_thres = 0; /* no VAD, record all */
/* set Julius default parameters for unspecified acoustic parameters */
apply_para(&(jconf->am_root->analysis.para), &(jconf->am_root->analysis.para_default));
/* set some values */
jconf->input.sfreq = jconf->am_root->analysis.para.smp_freq;
jconf->input.period = jconf->am_root->analysis.para.smp_period;
jconf->input.frameshift = jconf->am_root->analysis.para.frameshift;
jconf->input.framesize = jconf->am_root->analysis.para.framesize;
sfreq = jconf->am_root->analysis.para.smp_freq;
/* output file check */
if (!stout) {
if (access(filename, F_OK) == 0) {
if (access(filename, W_OK) == 0) {
fprintf(stderr, "Warning: overwriting file \"%s\"\n", filename);
} else {
perror("mkss");
return(1);
}
}
}
/* allocate speech store buffer */
samples = sfreq * slen / 1000;
speech = (SP16 *)mymalloc(sizeof(SP16) * samples);
/* allocate work area to compute spectrum */
wrk = WMP_work_new(&(jconf->am_root->analysis.para));
if (wrk == NULL) {
jlog("ERROR: m_fusion: failed to initialize MFCC computation for SS\n");
return -1;
}
/* initialize input device */
if (j_adin_init(recog) == FALSE) {
fprintf(stderr, "Error in initializing adin device\n");
return -1;
}
/* open device */
if (j_open_stream(recog, NULL) < 0) {
fprintf(stderr, "Error in opening adin device\n");
}
/* record mic input */
fprintf(stderr, "%dHz recording for %.2f seconds of noise\n", sfreq, (float)slen /(float)1000);
speechnum = 0;
adin_go(adin_callback, NULL, recog);
/* close device */
adin_end(recog->adin);
fprintf(stderr, "\n%d samples (%d bytes, %.1f sec) recorded\n", samples, samples * sizeof(SP16), (float)samples / (float)sfreq);
/* compute SS */
fprintf(stderr, "compute SS:\n");
fprintf(stderr, " fsize : %4d samples (%.1f msec)\n", jconf->input.framesize, (float)jconf->input.framesize * 1000.0/ (float)sfreq);
fprintf(stderr, " fshift: %4d samples (%.1f msec)\n", jconf->input.frameshift, (float)jconf->input.frameshift * 1000.0/ (float)sfreq);
ss = new_SS_calculate(speech, samples, &sslen, wrk, &(jconf->am_root->analysis.para));
fprintf(stderr, " points: %4d\n", sslen);
fprintf(stderr, "noise spectrum was measured\n");
/* open file for recording */
fprintf(stderr, "writing average noise spectrum to [%s]...", filename);
if (stout) {
fd = 1;
} else {
if ((fd = open(filename, O_CREAT | O_RDWR
#ifdef O_BINARY
| O_BINARY
#endif
, 0644)) == -1) {
perror("mkss");
return(1);
}
}
x = sslen;
#ifndef WORDS_BIGENDIAN
swap_bytes((char *)&x, sizeof(int), 1);
#endif
if (write(fd, &x, sizeof(int)) < sizeof(int)) {
perror("mkss");
return(1);
}
#ifndef WORDS_BIGENDIAN
swap_bytes((char *)ss, sizeof(float), sslen);
#endif
if (write(fd, ss, sslen * sizeof(float)) < sslen * sizeof(float)) {
perror("mkss");
return(1);
}
if (!stout) {
if (close(fd) < 0) {
perror("mkss");
return(1);
}
}
fprintf(stderr, "done\n");
WMP_free(wrk);
return 0;
}
/**
* <JA>
* 音声波形データから MFCC パラメータを抽出する.
* エンジンインスタンス内の MFCC 計算インスタンスごとにパラメータ抽出が
* 行われ，それぞれの mfcc->param に格納される.
*
* @param speech [in] 音声波形データ
* @param speechlen [in] @a speech の長さ（単位：サンプル数）
* @param recog [in] エンジンインスタンス
*
* @return 成功時 TRUE, エラー時 FALSE を返す.
* </JA>
* <EN>
* Extract MFCC parameters with sentence CMN from given waveform.
* Parameters will be computed for each MFCC calculation instance
* in the engine instance, and stored in mfcc->param for each.
*
* @param speech [in] buffer of speech waveform
* @param speechlen [in] length of @a speech in samples
* @param recog [in] engine instance
*
* @return TRUE on success, FALSE on error.
* </EN>
*
* @callgraph
* @callergraph
*/
boolean
wav2mfcc(SP16 speech[], int speechlen, Recog *recog)
{
int framenum;
int len;
Value *para;
MFCCCalc *mfcc;
int veclen;
int t, i;
/* calculate frame length from speech length, frame size and frame shift */
framenum = (int)((speechlen - recog->jconf->input.framesize) / recog->jconf->input.frameshift) + 1;
if (framenum < 1) {
jlog("WARNING: input too short (%d samples), ignored\n", speechlen);
return FALSE;
}
for(mfcc=recog->mfcclist;mfcc;mfcc=mfcc->next) {
if (mfcc->frontend.ssload_filename) {
/* setup for spectral subtraction using file */
if (mfcc->frontend.ssbuf == NULL) {
/* load noise spectrum for spectral subtraction from file (once) */
if ((mfcc->frontend.ssbuf = new_SS_load_from_file(mfcc->frontend.ssload_filename, &(mfcc->frontend.sslen))) == NULL) {
jlog("ERROR: wav2mfcc: failed to read noise spectrum from file \"%s\"\n", mfcc->frontend.ssload_filename);
return FALSE;
}
}
}
if (mfcc->frontend.sscalc) {
/* compute noise spectrum from head silence for each input */
len = mfcc->frontend.sscalc_len * recog->jconf->input.sfreq / 1000;
if (len > speechlen) len = speechlen;
#ifdef SSDEBUG
jlog("DEBUG: [%d]\n", len);
#endif
mfcc->frontend.ssbuf = new_SS_calculate(speech, len, &(mfcc->frontend.sslen), mfcc->frontend.mfccwrk_ss, mfcc->para);
}
}
/* compute mfcc from speech file for each mfcc instances */
for(mfcc=recog->mfcclist;mfcc;mfcc=mfcc->next) {
para = mfcc->para;
veclen = para->veclen * mfcc->splice;
if (framenum - (mfcc->splice - 1) < 1) {
jlog("WARNING: input too short (%d samples), ignored\n", speechlen);
return FALSE;
}
/* malloc new param */
param_init_content(mfcc->param);
if (param_alloc(mfcc->param, framenum, veclen) == FALSE) {
jlog("ERROR: failed to allocate memory for converted parameter vectors\n");
return FALSE;
}
if (mfcc->frontend.ssload_filename || mfcc->frontend.sscalc) {
/* make link from mfccs to this buffer */
mfcc->wrk->ssbuf = mfcc->frontend.ssbuf;
mfcc->wrk->ssbuflen = mfcc->frontend.sslen;
mfcc->wrk->ss_alpha = mfcc->frontend.ss_alpha;
mfcc->wrk->ss_floor = mfcc->frontend.ss_floor;
}
/* make MFCC from speech data */
if (Wav2MFCC(speech, mfcc->param->parvec, para, speechlen, mfcc->wrk, mfcc->cmn.wrk) == FALSE) {
jlog("ERROR: failed to compute features from input speech\n");
if (mfcc->frontend.sscalc) {
free(mfcc->frontend.ssbuf);
mfcc->frontend.ssbuf = NULL;
}
return FALSE;
}
/* splicing */
if (mfcc->splice > 1) {
for (t = 0; t < framenum - (mfcc->splice - 1); t++) {
for (i = 1; i < mfcc->splice; i++) {
memcpy(&(mfcc->param->parvec[t][para->veclen * i]), &(mfcc->param->parvec[t + i][0]), sizeof(VECT) * para->veclen);
}
}
}
/* set miscellaneous parameters */
mfcc->param->header.samplenum = framenum - (mfcc->splice - 1);
mfcc->param->header.wshift = para->smp_period * para->frameshift;
mfcc->param->header.sampsize = veclen * sizeof(VECT); /* not compressed */
mfcc->param->header.samptype = para->basetype;
if (para->delta) mfcc->param->header.samptype |= F_DELTA;
if (para->acc) mfcc->param->header.samptype |= F_ACCL;
if (para->energy) mfcc->param->header.samptype |= F_ENERGY;
if (para->c0) mfcc->param->header.samptype |= F_ZEROTH;
if (para->absesup) mfcc->param->header.samptype |= F_ENERGY_SUP;
if (para->cmn) mfcc->param->header.samptype |= F_CEPNORM;
mfcc->param->veclen = veclen;
mfcc->param->samplenum = framenum - (mfcc->splice - 1);
if (mfcc->frontend.sscalc) {
free(mfcc->frontend.ssbuf);
mfcc->frontend.ssbuf = NULL;
}
}
return TRUE;
}