void mfcc_calc( Mfcc *pMfcc, AUD_Int32s *Xabs, AUD_Int32s *mfcc )
{
AUD_Int32s lz, minLz;
AUD_Int32s i, j, k;
/* find the number of leading zeros in the largest number */
minLz = 32;
for ( k = 0; k < pMfcc->fftLen; k++ )
{
lz = clz( Xabs[k] );
if ( lz < minLz )
{
minLz = lz;
}
}
/* scale down the to have at least 15 leading zeros to prevent saturation in next step */
if ( minLz < 15 )
{
for ( k = 0; k < pMfcc->fftLen; k++ )
{
Xabs[k] = Xabs[k] >> (15 - minLz);
}
}
/* filter bank calculation */
melfb_calc( pMfcc->pFb, Xabs, pMfcc->fbBuffer );
if ( pMfcc->compressType == AUD_AMPCOMPRESS_LOG )
{
for ( k = 0; k < pMfcc->fbLen; k++ )
{
/* make sure we have no zeros in the ln() calulation */
if ( pMfcc->fbBuffer[k] <= 0 )
{
pMfcc->fbBuffer[k] = 1;
}
pMfcc->fbBuffer[k] = fp_ln( pMfcc->fbBuffer[k] );
}
}
else if ( pMfcc->compressType == AUD_AMPCOMPRESS_ROOT )
{
for ( k = 0; k < pMfcc->fbLen; k++ )
{
// in noisy environment, a bigger root lead to better feature discriminative capability
// refer to "Improving the noise-robustness of mel-frequency cepstral coefficients for speech recognition"
if ( pMfcc->fbBuffer[k] <= 0 )
{
pMfcc->fbBuffer[k] = 0;
}
else
{
pMfcc->fbBuffer[k] = (AUD_Int32s)round( pow( pMfcc->fbBuffer[k] / 65536.0, 0.14 ) * 65536.0 );
}
}
}
else
{
AUD_ASSERT( 0 );
}
/* DCT calculation */
dct_calc( pMfcc->pDct, pMfcc->fbBuffer, mfcc );
/* buffering */
for ( i = RAW_BUFLEN - 1; i >= 1; i-- )
{
for ( j = 0; j < pMfcc->mfccLen; j++ )
{
pMfcc->pState[i][j] = pMfcc->pState[i - 1][j];
}
}
for ( j = 0; j < pMfcc->mfccLen; j++ )
{
pMfcc->pState[0][j] = mfcc[j];
}
return;
}
int main(int argc, char **argv)
{
FFTComplex *tab, *tab1, *tab_ref;
FFTSample *tab2;
enum tf_transform transform = TRANSFORM_FFT;
FFTContext *m, *s;
#if FFT_FLOAT
RDFTContext *r;
DCTContext *d;
#endif /* FFT_FLOAT */
int it, i, err = 1;
int do_speed = 0, do_inverse = 0;
int fft_nbits = 9, fft_size;
double scale = 1.0;
AVLFG prng;
#if !AVFFT
s = av_mallocz(sizeof(*s));
m = av_mallocz(sizeof(*m));
#endif
#if !AVFFT && FFT_FLOAT
r = av_mallocz(sizeof(*r));
d = av_mallocz(sizeof(*d));
#endif
av_lfg_init(&prng, 1);
for (;;) {
int c = getopt(argc, argv, "hsimrdn:f:c:");
if (c == -1)
break;
switch (c) {
case 'h':
help();
return 1;
case 's':
do_speed = 1;
break;
case 'i':
do_inverse = 1;
break;
case 'm':
transform = TRANSFORM_MDCT;
break;
case 'r':
transform = TRANSFORM_RDFT;
break;
case 'd':
transform = TRANSFORM_DCT;
break;
case 'n':
fft_nbits = atoi(optarg);
break;
case 'f':
scale = atof(optarg);
break;
case 'c':
{
unsigned cpuflags = av_get_cpu_flags();
if (av_parse_cpu_caps(&cpuflags, optarg) < 0)
return 1;
av_force_cpu_flags(cpuflags);
break;
}
}
}
fft_size = 1 << fft_nbits;
tab = av_malloc_array(fft_size, sizeof(FFTComplex));
tab1 = av_malloc_array(fft_size, sizeof(FFTComplex));
tab_ref = av_malloc_array(fft_size, sizeof(FFTComplex));
tab2 = av_malloc_array(fft_size, sizeof(FFTSample));
if (!(tab && tab1 && tab_ref && tab2))
goto cleanup;
switch (transform) {
#if CONFIG_MDCT
case TRANSFORM_MDCT:
av_log(NULL, AV_LOG_INFO, "Scale factor is set to %f\n", scale);
if (do_inverse)
av_log(NULL, AV_LOG_INFO, "IMDCT");
else
av_log(NULL, AV_LOG_INFO, "MDCT");
mdct_init(&m, fft_nbits, do_inverse, scale);
break;
#endif /* CONFIG_MDCT */
case TRANSFORM_FFT:
if (do_inverse)
av_log(NULL, AV_LOG_INFO, "IFFT");
else
av_log(NULL, AV_LOG_INFO, "FFT");
fft_init(&s, fft_nbits, do_inverse);
if ((err = fft_ref_init(fft_nbits, do_inverse)) < 0)
goto cleanup;
break;
#if FFT_FLOAT
# if CONFIG_RDFT
case TRANSFORM_RDFT:
if (do_inverse)
av_log(NULL, AV_LOG_INFO, "IDFT_C2R");
else
av_log(NULL, AV_LOG_INFO, "DFT_R2C");
rdft_init(&r, fft_nbits, do_inverse ? IDFT_C2R : DFT_R2C);
if ((err = fft_ref_init(fft_nbits, do_inverse)) < 0)
goto cleanup;
break;
# endif /* CONFIG_RDFT */
# if CONFIG_DCT
case TRANSFORM_DCT:
if (do_inverse)
av_log(NULL, AV_LOG_INFO, "DCT_III");
else
av_log(NULL, AV_LOG_INFO, "DCT_II");
dct_init(&d, fft_nbits, do_inverse ? DCT_III : DCT_II);
break;
# endif /* CONFIG_DCT */
#endif /* FFT_FLOAT */
default:
av_log(NULL, AV_LOG_ERROR, "Requested transform not supported\n");
goto cleanup;
}
av_log(NULL, AV_LOG_INFO, " %d test\n", fft_size);
/* generate random data */
for (i = 0; i < fft_size; i++) {
tab1[i].re = frandom(&prng);
tab1[i].im = frandom(&prng);
}
/* checking result */
av_log(NULL, AV_LOG_INFO, "Checking...\n");
switch (transform) {
#if CONFIG_MDCT
case TRANSFORM_MDCT:
if (do_inverse) {
imdct_ref(&tab_ref->re, &tab1->re, fft_nbits);
imdct_calc(m, tab2, &tab1->re);
err = check_diff(&tab_ref->re, tab2, fft_size, scale);
} else {
mdct_ref(&tab_ref->re, &tab1->re, fft_nbits);
mdct_calc(m, tab2, &tab1->re);
err = check_diff(&tab_ref->re, tab2, fft_size / 2, scale);
}
break;
#endif /* CONFIG_MDCT */
case TRANSFORM_FFT:
memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
fft_permute(s, tab);
fft_calc(s, tab);
fft_ref(tab_ref, tab1, fft_nbits);
err = check_diff(&tab_ref->re, &tab->re, fft_size * 2, 1.0);
break;
#if FFT_FLOAT
#if CONFIG_RDFT
case TRANSFORM_RDFT:
{
int fft_size_2 = fft_size >> 1;
if (do_inverse) {
tab1[0].im = 0;
tab1[fft_size_2].im = 0;
for (i = 1; i < fft_size_2; i++) {
tab1[fft_size_2 + i].re = tab1[fft_size_2 - i].re;
tab1[fft_size_2 + i].im = -tab1[fft_size_2 - i].im;
}
memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
tab2[1] = tab1[fft_size_2].re;
rdft_calc(r, tab2);
fft_ref(tab_ref, tab1, fft_nbits);
for (i = 0; i < fft_size; i++) {
tab[i].re = tab2[i];
tab[i].im = 0;
}
err = check_diff(&tab_ref->re, &tab->re, fft_size * 2, 0.5);
} else {
for (i = 0; i < fft_size; i++) {
tab2[i] = tab1[i].re;
tab1[i].im = 0;
}
rdft_calc(r, tab2);
fft_ref(tab_ref, tab1, fft_nbits);
tab_ref[0].im = tab_ref[fft_size_2].re;
err = check_diff(&tab_ref->re, tab2, fft_size, 1.0);
}
break;
}
#endif /* CONFIG_RDFT */
#if CONFIG_DCT
case TRANSFORM_DCT:
memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
dct_calc(d, &tab->re);
if (do_inverse)
idct_ref(&tab_ref->re, &tab1->re, fft_nbits);
else
dct_ref(&tab_ref->re, &tab1->re, fft_nbits);
err = check_diff(&tab_ref->re, &tab->re, fft_size, 1.0);
break;
#endif /* CONFIG_DCT */
#endif /* FFT_FLOAT */
}
/* do a speed test */
if (do_speed) {
int64_t time_start, duration;
int nb_its;
av_log(NULL, AV_LOG_INFO, "Speed test...\n");
/* we measure during about 1 seconds */
nb_its = 1;
for (;;) {
time_start = av_gettime_relative();
for (it = 0; it < nb_its; it++) {
switch (transform) {
case TRANSFORM_MDCT:
if (do_inverse)
imdct_calc(m, &tab->re, &tab1->re);
else
mdct_calc(m, &tab->re, &tab1->re);
break;
case TRANSFORM_FFT:
memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
fft_calc(s, tab);
break;
#if FFT_FLOAT
case TRANSFORM_RDFT:
memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
rdft_calc(r, tab2);
break;
case TRANSFORM_DCT:
memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
dct_calc(d, tab2);
break;
#endif /* FFT_FLOAT */
}
}
duration = av_gettime_relative() - time_start;
if (duration >= 1000000)
break;
nb_its *= 2;
}
av_log(NULL, AV_LOG_INFO,
"time: %0.1f us/transform [total time=%0.2f s its=%d]\n",
(double) duration / nb_its,
(double) duration / 1000000.0,
nb_its);
}
switch (transform) {
#if CONFIG_MDCT
case TRANSFORM_MDCT:
mdct_end(m);
break;
#endif /* CONFIG_MDCT */
case TRANSFORM_FFT:
fft_end(s);
break;
#if FFT_FLOAT
# if CONFIG_RDFT
case TRANSFORM_RDFT:
rdft_end(r);
break;
# endif /* CONFIG_RDFT */
# if CONFIG_DCT
case TRANSFORM_DCT:
dct_end(d);
break;
# endif /* CONFIG_DCT */
#endif /* FFT_FLOAT */
}
cleanup:
av_free(tab);
av_free(tab1);
av_free(tab2);
av_free(tab_ref);
av_free(exptab);
#if !AVFFT
av_free(s);
av_free(m);
#endif
#if !AVFFT && FFT_FLOAT
av_free(r);
av_free(d);
#endif
if (err)
printf("Error: %d.\n", err);
return !!err;
}
