void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info_out, const CELTMode *celt_mode, const void *x, int len, int offset, int c1, int c2, int C, int lsb_depth, downmix_func downmix)
{
int i, b;
const kiss_fft_state *kfft;
VARDECL(kiss_fft_cpx, in);
VARDECL(kiss_fft_cpx, out);
int N = 480, N2=240;
float * OPUS_RESTRICT A = tonal->angle;
float * OPUS_RESTRICT dA = tonal->d_angle;
float * OPUS_RESTRICT d2A = tonal->d2_angle;
VARDECL(float, tonality);
VARDECL(float, noisiness);
float band_tonality[NB_TBANDS];
float logE[NB_TBANDS];
float BFCC[8];
float features[25];
float frame_tonality;
float max_frame_tonality;
/*float tw_sum=0;*/
float frame_noisiness;
const float pi4 = (float)(M_PI*M_PI*M_PI*M_PI);
float slope=0;
float frame_stationarity;
float relativeE;
float frame_probs[2];
float alpha, alphaE, alphaE2;
float frame_loudness;
float bandwidth_mask;
int bandwidth=0;
float maxE = 0;
float noise_floor;
int remaining;
AnalysisInfo *info;
SAVE_STACK;
tonal->last_transition++;
alpha = 1.f/IMIN(20, 1+tonal->count);
alphaE = 1.f/IMIN(50, 1+tonal->count);
alphaE2 = 1.f/IMIN(1000, 1+tonal->count);
if (tonal->count<4)
tonal->music_prob = .5;
kfft = celt_mode->mdct.kfft[0];
if (tonal->count==0)
tonal->mem_fill = 240;
downmix(x, &tonal->inmem[tonal->mem_fill], IMIN(len, ANALYSIS_BUF_SIZE-tonal->mem_fill), offset, c1, c2, C);
if (tonal->mem_fill+len < ANALYSIS_BUF_SIZE)
{
tonal->mem_fill += len;
/* Don't have enough to update the analysis */
RESTORE_STACK;
return;
}
info = &tonal->info[tonal->write_pos++];
if (tonal->write_pos>=DETECT_SIZE)
tonal->write_pos-=DETECT_SIZE;
ALLOC(in, 480, kiss_fft_cpx);
ALLOC(out, 480, kiss_fft_cpx);
ALLOC(tonality, 240, float);
ALLOC(noisiness, 240, float);
for (i=0;i<N2;i++)
{
float w = analysis_window[i];
in[i].r = (kiss_fft_scalar)(w*tonal->inmem[i]);
in[i].i = (kiss_fft_scalar)(w*tonal->inmem[N2+i]);
in[N-i-1].r = (kiss_fft_scalar)(w*tonal->inmem[N-i-1]);
in[N-i-1].i = (kiss_fft_scalar)(w*tonal->inmem[N+N2-i-1]);
}
OPUS_MOVE(tonal->inmem, tonal->inmem+ANALYSIS_BUF_SIZE-240, 240);
remaining = len - (ANALYSIS_BUF_SIZE-tonal->mem_fill);
downmix(x, &tonal->inmem[240], remaining, offset+ANALYSIS_BUF_SIZE-tonal->mem_fill, c1, c2, C);
tonal->mem_fill = 240 + remaining;
opus_fft(kfft, in, out);
for (i=1;i<N2;i++)
{
float X1r, X2r, X1i, X2i;
float angle, d_angle, d2_angle;
float angle2, d_angle2, d2_angle2;
float mod1, mod2, avg_mod;
X1r = (float)out[i].r+out[N-i].r;
X1i = (float)out[i].i-out[N-i].i;
X2r = (float)out[i].i+out[N-i].i;
X2i = (float)out[N-i].r-out[i].r;
angle = (float)(.5f/M_PI)*fast_atan2f(X1i, X1r);
d_angle = angle - A[i];
d2_angle = d_angle - dA[i];
angle2 = (float)(.5f/M_PI)*fast_atan2f(X2i, X2r);
d_angle2 = angle2 - angle;
d2_angle2 = d_angle2 - d_angle;
mod1 = d2_angle - (float)floor(.5+d2_angle);
noisiness[i] = ABS16(mod1);
mod1 *= mod1;
mod1 *= mod1;
mod2 = d2_angle2 - (float)floor(.5+d2_angle2);
noisiness[i] += ABS16(mod2);
mod2 *= mod2;
mod2 *= mod2;
avg_mod = .25f*(d2A[i]+2.f*mod1+mod2);
tonality[i] = 1.f/(1.f+40.f*16.f*pi4*avg_mod)-.015f;
A[i] = angle2;
dA[i] = d_angle2;
d2A[i] = mod2;
}
frame_tonality = 0;
max_frame_tonality = 0;
/*tw_sum = 0;*/
info->activity = 0;
frame_noisiness = 0;
frame_stationarity = 0;
if (!tonal->count)
{
for (b=0;b<NB_TBANDS;b++)
{
tonal->lowE[b] = 1e10;
tonal->highE[b] = -1e10;
}
}
relativeE = 0;
frame_loudness = 0;
for (b=0;b<NB_TBANDS;b++)
{
float E=0, tE=0, nE=0;
float L1, L2;
float stationarity;
for (i=tbands[b];i<tbands[b+1];i++)
{
float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
+ out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
#ifdef FIXED_POINT
/* FIXME: It's probably best to change the BFCC filter initial state instead */
binE *= 5.55e-17f;
#endif
E += binE;
tE += binE*tonality[i];
nE += binE*2.f*(.5f-noisiness[i]);
}
tonal->E[tonal->E_count][b] = E;
frame_noisiness += nE/(1e-15f+E);
frame_loudness += (float)sqrt(E+1e-10f);
logE[b] = (float)log(E+1e-10f);
tonal->lowE[b] = MIN32(logE[b], tonal->lowE[b]+.01f);
tonal->highE[b] = MAX32(logE[b], tonal->highE[b]-.1f);
if (tonal->highE[b] < tonal->lowE[b]+1.f)
{
tonal->highE[b]+=.5f;
tonal->lowE[b]-=.5f;
}
relativeE += (logE[b]-tonal->lowE[b])/(1e-15f+tonal->highE[b]-tonal->lowE[b]);
L1=L2=0;
for (i=0;i<NB_FRAMES;i++)
{
L1 += (float)sqrt(tonal->E[i][b]);
L2 += tonal->E[i][b];
}
stationarity = MIN16(0.99f,L1/(float)sqrt(1e-15+NB_FRAMES*L2));
stationarity *= stationarity;
stationarity *= stationarity;
frame_stationarity += stationarity;
/*band_tonality[b] = tE/(1e-15+E)*/;
band_tonality[b] = MAX16(tE/(1e-15f+E), stationarity*tonal->prev_band_tonality[b]);
#if 0
if (b>=NB_TONAL_SKIP_BANDS)
{
frame_tonality += tweight[b]*band_tonality[b];
tw_sum += tweight[b];
}
#else
frame_tonality += band_tonality[b];
if (b>=NB_TBANDS-NB_TONAL_SKIP_BANDS)
frame_tonality -= band_tonality[b-NB_TBANDS+NB_TONAL_SKIP_BANDS];
#endif
max_frame_tonality = MAX16(max_frame_tonality, (1.f+.03f*(b-NB_TBANDS))*frame_tonality);
slope += band_tonality[b]*(b-8);
/*printf("%f %f ", band_tonality[b], stationarity);*/
tonal->prev_band_tonality[b] = band_tonality[b];
}
bandwidth_mask = 0;
bandwidth = 0;
maxE = 0;
noise_floor = 5.7e-4f/(1<<(IMAX(0,lsb_depth-8)));
#ifdef FIXED_POINT
noise_floor *= 1<<(15+SIG_SHIFT);
#endif
noise_floor *= noise_floor;
for (b=0;b<NB_TOT_BANDS;b++)
{
float E=0;
int band_start, band_end;
/* Keep a margin of 300 Hz for aliasing */
band_start = extra_bands[b];
band_end = extra_bands[b+1];
for (i=band_start;i<band_end;i++)
{
float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
+ out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
E += binE;
}
maxE = MAX32(maxE, E);
tonal->meanE[b] = MAX32((1-alphaE2)*tonal->meanE[b], E);
E = MAX32(E, tonal->meanE[b]);
/* Use a simple follower with 13 dB/Bark slope for spreading function */
bandwidth_mask = MAX32(.05f*bandwidth_mask, E);
/* Consider the band "active" only if all these conditions are met:
1) less than 10 dB below the simple follower
2) less than 90 dB below the peak band (maximal masking possible considering
both the ATH and the loudness-dependent slope of the spreading function)
3) above the PCM quantization noise floor
*/
if (E>.1*bandwidth_mask && E*1e9f > maxE && E > noise_floor*(band_end-band_start))
bandwidth = b;
}
if (tonal->count<=2)
bandwidth = 20;
frame_loudness = 20*(float)log10(frame_loudness);
tonal->Etracker = MAX32(tonal->Etracker-.03f, frame_loudness);
tonal->lowECount *= (1-alphaE);
if (frame_loudness < tonal->Etracker-30)
tonal->lowECount += alphaE;
for (i=0;i<8;i++)
{
float sum=0;
for (b=0;b<16;b++)
sum += dct_table[i*16+b]*logE[b];
BFCC[i] = sum;
}
frame_stationarity /= NB_TBANDS;
relativeE /= NB_TBANDS;
if (tonal->count<10)
relativeE = .5;
frame_noisiness /= NB_TBANDS;
#if 1
info->activity = frame_noisiness + (1-frame_noisiness)*relativeE;
#else
info->activity = .5*(1+frame_noisiness-frame_stationarity);
#endif
frame_tonality = (max_frame_tonality/(NB_TBANDS-NB_TONAL_SKIP_BANDS));
frame_tonality = MAX16(frame_tonality, tonal->prev_tonality*.8f);
tonal->prev_tonality = frame_tonality;
slope /= 8*8;
info->tonality_slope = slope;
tonal->E_count = (tonal->E_count+1)%NB_FRAMES;
tonal->count++;
info->tonality = frame_tonality;
for (i=0;i<4;i++)
features[i] = -0.12299f*(BFCC[i]+tonal->mem[i+24]) + 0.49195f*(tonal->mem[i]+tonal->mem[i+16]) + 0.69693f*tonal->mem[i+8] - 1.4349f*tonal->cmean[i];
for (i=0;i<4;i++)
tonal->cmean[i] = (1-alpha)*tonal->cmean[i] + alpha*BFCC[i];
for (i=0;i<4;i++)
features[4+i] = 0.63246f*(BFCC[i]-tonal->mem[i+24]) + 0.31623f*(tonal->mem[i]-tonal->mem[i+16]);
for (i=0;i<3;i++)
features[8+i] = 0.53452f*(BFCC[i]+tonal->mem[i+24]) - 0.26726f*(tonal->mem[i]+tonal->mem[i+16]) -0.53452f*tonal->mem[i+8];
if (tonal->count > 5)
{
for (i=0;i<9;i++)
tonal->std[i] = (1-alpha)*tonal->std[i] + alpha*features[i]*features[i];
}
for (i=0;i<8;i++)
{
tonal->mem[i+24] = tonal->mem[i+16];
tonal->mem[i+16] = tonal->mem[i+8];
tonal->mem[i+8] = tonal->mem[i];
tonal->mem[i] = BFCC[i];
}
for (i=0;i<9;i++)
features[11+i] = (float)sqrt(tonal->std[i]);
features[20] = info->tonality;
features[21] = info->activity;
features[22] = frame_stationarity;
features[23] = info->tonality_slope;
features[24] = tonal->lowECount;
#ifndef DISABLE_FLOAT_API
mlp_process(&net, features, frame_probs);
frame_probs[0] = .5f*(frame_probs[0]+1);
/* Curve fitting between the MLP probability and the actual probability */
frame_probs[0] = .01f + 1.21f*frame_probs[0]*frame_probs[0] - .23f*(float)pow(frame_probs[0], 10);
/* Probability of active audio (as opposed to silence) */
frame_probs[1] = .5f*frame_probs[1]+.5f;
/* Consider that silence has a 50-50 probability. */
frame_probs[0] = frame_probs[1]*frame_probs[0] + (1-frame_probs[1])*.5f;
/*printf("%f %f ", frame_probs[0], frame_probs[1]);*/
{
/* Probability of state transition */
float tau;
/* Represents independence of the MLP probabilities, where
beta=1 means fully independent. */
float beta;
/* Denormalized probability of speech (p0) and music (p1) after update */
float p0, p1;
/* Probabilities for "all speech" and "all music" */
float s0, m0;
/* Probability sum for renormalisation */
float psum;
/* Instantaneous probability of speech and music, with beta pre-applied. */
float speech0;
float music0;
/* One transition every 3 minutes of active audio */
tau = .00005f*frame_probs[1];
beta = .05f;
if (1) {
/* Adapt beta based on how "unexpected" the new prob is */
float p, q;
p = MAX16(.05f,MIN16(.95f,frame_probs[0]));
q = MAX16(.05f,MIN16(.95f,tonal->music_prob));
beta = .01f+.05f*ABS16(p-q)/(p*(1-q)+q*(1-p));
}
/* p0 and p1 are the probabilities of speech and music at this frame
using only information from previous frame and applying the
state transition model */
p0 = (1-tonal->music_prob)*(1-tau) + tonal->music_prob *tau;
p1 = tonal->music_prob *(1-tau) + (1-tonal->music_prob)*tau;
/* We apply the current probability with exponent beta to work around
the fact that the probability estimates aren't independent. */
p0 *= (float)pow(1-frame_probs[0], beta);
p1 *= (float)pow(frame_probs[0], beta);
/* Normalise the probabilities to get the Marokv probability of music. */
tonal->music_prob = p1/(p0+p1);
info->music_prob = tonal->music_prob;
/* This chunk of code deals with delayed decision. */
psum=1e-20f;
/* Instantaneous probability of speech and music, with beta pre-applied. */
speech0 = (float)pow(1-frame_probs[0], beta);
music0 = (float)pow(frame_probs[0], beta);
if (tonal->count==1)
{
tonal->pspeech[0]=.5;
tonal->pmusic [0]=.5;
}
/* Updated probability of having only speech (s0) or only music (m0),
before considering the new observation. */
s0 = tonal->pspeech[0] + tonal->pspeech[1];
m0 = tonal->pmusic [0] + tonal->pmusic [1];
/* Updates s0 and m0 with instantaneous probability. */
tonal->pspeech[0] = s0*(1-tau)*speech0;
tonal->pmusic [0] = m0*(1-tau)*music0;
/* Propagate the transition probabilities */
for (i=1;i<DETECT_SIZE-1;i++)
{
tonal->pspeech[i] = tonal->pspeech[i+1]*speech0;
tonal->pmusic [i] = tonal->pmusic [i+1]*music0;
}
/* Probability that the latest frame is speech, when all the previous ones were music. */
tonal->pspeech[DETECT_SIZE-1] = m0*tau*speech0;
/* Probability that the latest frame is music, when all the previous ones were speech. */
tonal->pmusic [DETECT_SIZE-1] = s0*tau*music0;
/* Renormalise probabilities to 1 */
for (i=0;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i] + tonal->pmusic[i];
psum = 1.f/psum;
for (i=0;i<DETECT_SIZE;i++)
{
tonal->pspeech[i] *= psum;
tonal->pmusic [i] *= psum;
}
psum = tonal->pmusic[0];
for (i=1;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i];
/* Estimate our confidence in the speech/music decisions */
if (frame_probs[1]>.75)
{
if (tonal->music_prob>.9)
{
float adapt;
adapt = 1.f/(++tonal->music_confidence_count);
tonal->music_confidence_count = IMIN(tonal->music_confidence_count, 500);
tonal->music_confidence += adapt*MAX16(-.2f,frame_probs[0]-tonal->music_confidence);
}
if (tonal->music_prob<.1)
{
float adapt;
adapt = 1.f/(++tonal->speech_confidence_count);
tonal->speech_confidence_count = IMIN(tonal->speech_confidence_count, 500);
tonal->speech_confidence += adapt*MIN16(.2f,frame_probs[0]-tonal->speech_confidence);
}
} else {
if (tonal->music_confidence_count==0)
tonal->music_confidence = .9f;
if (tonal->speech_confidence_count==0)
tonal->speech_confidence = .1f;
}
}
if (tonal->last_music != (tonal->music_prob>.5f))
tonal->last_transition=0;
tonal->last_music = tonal->music_prob>.5f;
#else
info->music_prob = 0;
#endif
/*for (i=0;i<25;i++)
printf("%f ", features[i]);
printf("\n");*/
info->bandwidth = bandwidth;
/*printf("%d %d\n", info->bandwidth, info->opus_bandwidth);*/
info->noisiness = frame_noisiness;
info->valid = 1;
if (info_out!=NULL)
OPUS_COPY(info_out, info, 1);
RESTORE_STACK;
}
static void tonality_analysis(TonalityAnalysisState *tonal, const CELTMode *celt_mode, const void *x, int len, int offset, int c1, int c2, int C, int lsb_depth, downmix_func downmix)
{
int i, b;
const kiss_fft_state *kfft;
VARDECL(kiss_fft_cpx, in);
VARDECL(kiss_fft_cpx, out);
int N = 480, N2=240;
float * OPUS_RESTRICT A = tonal->angle;
float * OPUS_RESTRICT dA = tonal->d_angle;
float * OPUS_RESTRICT d2A = tonal->d2_angle;
VARDECL(float, tonality);
VARDECL(float, noisiness);
float band_tonality[NB_TBANDS];
float logE[NB_TBANDS];
float BFCC[8];
float features[25];
float frame_tonality;
float max_frame_tonality;
/*float tw_sum=0;*/
float frame_noisiness;
const float pi4 = (float)(M_PI*M_PI*M_PI*M_PI);
float slope=0;
float frame_stationarity;
float relativeE;
float frame_probs[2];
float alpha, alphaE, alphaE2;
float frame_loudness;
float bandwidth_mask;
int bandwidth=0;
float maxE = 0;
float noise_floor;
int remaining;
AnalysisInfo *info;
float hp_ener;
float tonality2[240];
float midE[8];
float spec_variability=0;
float band_log2[NB_TBANDS+1];
float leakage_from[NB_TBANDS+1];
float leakage_to[NB_TBANDS+1];
SAVE_STACK;
alpha = 1.f/IMIN(10, 1+tonal->count);
alphaE = 1.f/IMIN(25, 1+tonal->count);
alphaE2 = 1.f/IMIN(500, 1+tonal->count);
if (tonal->Fs == 48000)
{
/* len and offset are now at 24 kHz. */
len/= 2;
offset /= 2;
} else if (tonal->Fs == 16000) {
len = 3*len/2;
offset = 3*offset/2;
}
if (tonal->count<4) {
if (tonal->application == OPUS_APPLICATION_VOIP)
tonal->music_prob = .1f;
else
tonal->music_prob = .625f;
}
kfft = celt_mode->mdct.kfft[0];
if (tonal->count==0)
tonal->mem_fill = 240;
tonal->hp_ener_accum += (float)downmix_and_resample(downmix, x,
&tonal->inmem[tonal->mem_fill], tonal->downmix_state,
IMIN(len, ANALYSIS_BUF_SIZE-tonal->mem_fill), offset, c1, c2, C, tonal->Fs);
if (tonal->mem_fill+len < ANALYSIS_BUF_SIZE)
{
tonal->mem_fill += len;
/* Don't have enough to update the analysis */
RESTORE_STACK;
return;
}
hp_ener = tonal->hp_ener_accum;
info = &tonal->info[tonal->write_pos++];
if (tonal->write_pos>=DETECT_SIZE)
tonal->write_pos-=DETECT_SIZE;
ALLOC(in, 480, kiss_fft_cpx);
ALLOC(out, 480, kiss_fft_cpx);
ALLOC(tonality, 240, float);
ALLOC(noisiness, 240, float);
for (i=0;i<N2;i++)
{
float w = analysis_window[i];
in[i].r = (kiss_fft_scalar)(w*tonal->inmem[i]);
in[i].i = (kiss_fft_scalar)(w*tonal->inmem[N2+i]);
in[N-i-1].r = (kiss_fft_scalar)(w*tonal->inmem[N-i-1]);
in[N-i-1].i = (kiss_fft_scalar)(w*tonal->inmem[N+N2-i-1]);
}
OPUS_MOVE(tonal->inmem, tonal->inmem+ANALYSIS_BUF_SIZE-240, 240);
remaining = len - (ANALYSIS_BUF_SIZE-tonal->mem_fill);
tonal->hp_ener_accum = (float)downmix_and_resample(downmix, x,
&tonal->inmem[240], tonal->downmix_state, remaining,
offset+ANALYSIS_BUF_SIZE-tonal->mem_fill, c1, c2, C, tonal->Fs);
tonal->mem_fill = 240 + remaining;
opus_fft(kfft, in, out, tonal->arch);
#ifndef FIXED_POINT
/* If there's any NaN on the input, the entire output will be NaN, so we only need to check one value. */
if (celt_isnan(out[0].r))
{
info->valid = 0;
RESTORE_STACK;
return;
}
#endif
for (i=1;i<N2;i++)
{
float X1r, X2r, X1i, X2i;
float angle, d_angle, d2_angle;
float angle2, d_angle2, d2_angle2;
float mod1, mod2, avg_mod;
X1r = (float)out[i].r+out[N-i].r;
X1i = (float)out[i].i-out[N-i].i;
X2r = (float)out[i].i+out[N-i].i;
X2i = (float)out[N-i].r-out[i].r;
angle = (float)(.5f/M_PI)*fast_atan2f(X1i, X1r);
d_angle = angle - A[i];
d2_angle = d_angle - dA[i];
angle2 = (float)(.5f/M_PI)*fast_atan2f(X2i, X2r);
d_angle2 = angle2 - angle;
d2_angle2 = d_angle2 - d_angle;
mod1 = d2_angle - (float)float2int(d2_angle);
noisiness[i] = ABS16(mod1);
mod1 *= mod1;
mod1 *= mod1;
mod2 = d2_angle2 - (float)float2int(d2_angle2);
noisiness[i] += ABS16(mod2);
mod2 *= mod2;
mod2 *= mod2;
avg_mod = .25f*(d2A[i]+mod1+2*mod2);
/* This introduces an extra delay of 2 frames in the detection. */
tonality[i] = 1.f/(1.f+40.f*16.f*pi4*avg_mod)-.015f;
/* No delay on this detection, but it's less reliable. */
tonality2[i] = 1.f/(1.f+40.f*16.f*pi4*mod2)-.015f;
A[i] = angle2;
dA[i] = d_angle2;
d2A[i] = mod2;
}
for (i=2;i<N2-1;i++)
{
float tt = MIN32(tonality2[i], MAX32(tonality2[i-1], tonality2[i+1]));
tonality[i] = .9f*MAX32(tonality[i], tt-.1f);
}
frame_tonality = 0;
max_frame_tonality = 0;
/*tw_sum = 0;*/
info->activity = 0;
frame_noisiness = 0;
frame_stationarity = 0;
if (!tonal->count)
{
for (b=0;b<NB_TBANDS;b++)
{
tonal->lowE[b] = 1e10;
tonal->highE[b] = -1e10;
}
}
relativeE = 0;
frame_loudness = 0;
/* The energy of the very first band is special because of DC. */
{
float E = 0;
float X1r, X2r;
X1r = 2*(float)out[0].r;
X2r = 2*(float)out[0].i;
E = X1r*X1r + X2r*X2r;
for (i=1;i<4;i++)
{
float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
+ out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
E += binE;
}
E = SCALE_ENER(E);
band_log2[0] = .5f*1.442695f*(float)log(E+1e-10f);
}
for (b=0;b<NB_TBANDS;b++)
{
float E=0, tE=0, nE=0;
float L1, L2;
float stationarity;
for (i=tbands[b];i<tbands[b+1];i++)
{
float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
+ out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
binE = SCALE_ENER(binE);
E += binE;
tE += binE*MAX32(0, tonality[i]);
nE += binE*2.f*(.5f-noisiness[i]);
}
#ifndef FIXED_POINT
/* Check for extreme band energies that could cause NaNs later. */
if (!(E<1e9f) || celt_isnan(E))
{
info->valid = 0;
RESTORE_STACK;
return;
}
#endif
tonal->E[tonal->E_count][b] = E;
frame_noisiness += nE/(1e-15f+E);
frame_loudness += (float)sqrt(E+1e-10f);
logE[b] = (float)log(E+1e-10f);
band_log2[b+1] = .5f*1.442695f*(float)log(E+1e-10f);
tonal->logE[tonal->E_count][b] = logE[b];
if (tonal->count==0)
tonal->highE[b] = tonal->lowE[b] = logE[b];
if (tonal->highE[b] > tonal->lowE[b] + 7.5)
{
if (tonal->highE[b] - logE[b] > logE[b] - tonal->lowE[b])
tonal->highE[b] -= .01f;
else
tonal->lowE[b] += .01f;
}
if (logE[b] > tonal->highE[b])
{
tonal->highE[b] = logE[b];
tonal->lowE[b] = MAX32(tonal->highE[b]-15, tonal->lowE[b]);
} else if (logE[b] < tonal->lowE[b])
{
tonal->lowE[b] = logE[b];
tonal->highE[b] = MIN32(tonal->lowE[b]+15, tonal->highE[b]);
}
relativeE += (logE[b]-tonal->lowE[b])/(1e-15f + (tonal->highE[b]-tonal->lowE[b]));
L1=L2=0;
for (i=0;i<NB_FRAMES;i++)
{
L1 += (float)sqrt(tonal->E[i][b]);
L2 += tonal->E[i][b];
}
stationarity = MIN16(0.99f,L1/(float)sqrt(1e-15+NB_FRAMES*L2));
stationarity *= stationarity;
stationarity *= stationarity;
frame_stationarity += stationarity;
/*band_tonality[b] = tE/(1e-15+E)*/;
band_tonality[b] = MAX16(tE/(1e-15f+E), stationarity*tonal->prev_band_tonality[b]);
#if 0
if (b>=NB_TONAL_SKIP_BANDS)
{
frame_tonality += tweight[b]*band_tonality[b];
tw_sum += tweight[b];
}
#else
frame_tonality += band_tonality[b];
if (b>=NB_TBANDS-NB_TONAL_SKIP_BANDS)
frame_tonality -= band_tonality[b-NB_TBANDS+NB_TONAL_SKIP_BANDS];
#endif
max_frame_tonality = MAX16(max_frame_tonality, (1.f+.03f*(b-NB_TBANDS))*frame_tonality);
slope += band_tonality[b]*(b-8);
/*printf("%f %f ", band_tonality[b], stationarity);*/
tonal->prev_band_tonality[b] = band_tonality[b];
}
leakage_from[0] = band_log2[0];
leakage_to[0] = band_log2[0] - LEAKAGE_OFFSET;
for (b=1;b<NB_TBANDS+1;b++)
{
float leak_slope = LEAKAGE_SLOPE*(tbands[b]-tbands[b-1])/4;
leakage_from[b] = MIN16(leakage_from[b-1]+leak_slope, band_log2[b]);
leakage_to[b] = MAX16(leakage_to[b-1]-leak_slope, band_log2[b]-LEAKAGE_OFFSET);
}
for (b=NB_TBANDS-2;b>=0;b--)
{
float leak_slope = LEAKAGE_SLOPE*(tbands[b+1]-tbands[b])/4;
leakage_from[b] = MIN16(leakage_from[b+1]+leak_slope, leakage_from[b]);
leakage_to[b] = MAX16(leakage_to[b+1]-leak_slope, leakage_to[b]);
}
celt_assert(NB_TBANDS+1 <= LEAK_BANDS);
for (b=0;b<NB_TBANDS+1;b++)
{
/* leak_boost[] is made up of two terms. The first, based on leakage_to[],
represents the boost needed to overcome the amount of analysis leakage
cause in a weaker band b by louder neighbouring bands.
The second, based on leakage_from[], applies to a loud band b for
which the quantization noise causes synthesis leakage to the weaker
neighbouring bands. */
float boost = MAX16(0, leakage_to[b] - band_log2[b]) +
MAX16(0, band_log2[b] - (leakage_from[b]+LEAKAGE_OFFSET));
info->leak_boost[b] = IMIN(255, (int)floor(.5 + 64.f*boost));
}
for (;b<LEAK_BANDS;b++) info->leak_boost[b] = 0;
for (i=0;i<NB_FRAMES;i++)
{
int j;
float mindist = 1e15f;
for (j=0;j<NB_FRAMES;j++)
{
int k;
float dist=0;
for (k=0;k<NB_TBANDS;k++)
{
float tmp;
tmp = tonal->logE[i][k] - tonal->logE[j][k];
dist += tmp*tmp;
}
if (j!=i)
mindist = MIN32(mindist, dist);
}
spec_variability += mindist;
}
spec_variability = (float)sqrt(spec_variability/NB_FRAMES/NB_TBANDS);
bandwidth_mask = 0;
bandwidth = 0;
maxE = 0;
noise_floor = 5.7e-4f/(1<<(IMAX(0,lsb_depth-8)));
noise_floor *= noise_floor;
for (b=0;b<NB_TBANDS;b++)
{
float E=0;
int band_start, band_end;
/* Keep a margin of 300 Hz for aliasing */
band_start = tbands[b];
band_end = tbands[b+1];
for (i=band_start;i<band_end;i++)
{
float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
+ out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
E += binE;
}
E = SCALE_ENER(E);
maxE = MAX32(maxE, E);
tonal->meanE[b] = MAX32((1-alphaE2)*tonal->meanE[b], E);
E = MAX32(E, tonal->meanE[b]);
/* Use a simple follower with 13 dB/Bark slope for spreading function */
bandwidth_mask = MAX32(.05f*bandwidth_mask, E);
/* Consider the band "active" only if all these conditions are met:
1) less than 10 dB below the simple follower
2) less than 90 dB below the peak band (maximal masking possible considering
both the ATH and the loudness-dependent slope of the spreading function)
3) above the PCM quantization noise floor
We use b+1 because the first CELT band isn't included in tbands[]
*/
if (E>.1*bandwidth_mask && E*1e9f > maxE && E > noise_floor*(band_end-band_start))
bandwidth = b+1;
}
/* Special case for the last two bands, for which we don't have spectrum but only
the energy above 12 kHz. */
if (tonal->Fs == 48000) {
float ratio;
float E = hp_ener*(1.f/(240*240));
ratio = tonal->prev_bandwidth==20 ? 0.03f : 0.07f;
#ifdef FIXED_POINT
/* silk_resampler_down2_hp() shifted right by an extra 8 bits. */
E *= 256.f*(1.f/Q15ONE)*(1.f/Q15ONE);
#endif
maxE = MAX32(maxE, E);
tonal->meanE[b] = MAX32((1-alphaE2)*tonal->meanE[b], E);
E = MAX32(E, tonal->meanE[b]);
/* Use a simple follower with 13 dB/Bark slope for spreading function */
bandwidth_mask = MAX32(.05f*bandwidth_mask, E);
if (E>ratio*bandwidth_mask && E*1e9f > maxE && E > noise_floor*160)
bandwidth = 20;
/* This detector is unreliable, so if the bandwidth is close to SWB, assume it's FB. */
if (bandwidth >= 17)
bandwidth = 20;
}
if (tonal->count<=2)
bandwidth = 20;
frame_loudness = 20*(float)log10(frame_loudness);
tonal->Etracker = MAX32(tonal->Etracker-.003f, frame_loudness);
tonal->lowECount *= (1-alphaE);
if (frame_loudness < tonal->Etracker-30)
tonal->lowECount += alphaE;
for (i=0;i<8;i++)
{
float sum=0;
for (b=0;b<16;b++)
sum += dct_table[i*16+b]*logE[b];
BFCC[i] = sum;
}
for (i=0;i<8;i++)
{
float sum=0;
for (b=0;b<16;b++)
sum += dct_table[i*16+b]*.5f*(tonal->highE[b]+tonal->lowE[b]);
midE[i] = sum;
}
frame_stationarity /= NB_TBANDS;
relativeE /= NB_TBANDS;
if (tonal->count<10)
relativeE = .5f;
frame_noisiness /= NB_TBANDS;
#if 1
info->activity = frame_noisiness + (1-frame_noisiness)*relativeE;
#else
info->activity = .5*(1+frame_noisiness-frame_stationarity);
#endif
frame_tonality = (max_frame_tonality/(NB_TBANDS-NB_TONAL_SKIP_BANDS));
frame_tonality = MAX16(frame_tonality, tonal->prev_tonality*.8f);
tonal->prev_tonality = frame_tonality;
slope /= 8*8;
info->tonality_slope = slope;
tonal->E_count = (tonal->E_count+1)%NB_FRAMES;
tonal->count = IMIN(tonal->count+1, ANALYSIS_COUNT_MAX);
info->tonality = frame_tonality;
for (i=0;i<4;i++)
features[i] = -0.12299f*(BFCC[i]+tonal->mem[i+24]) + 0.49195f*(tonal->mem[i]+tonal->mem[i+16]) + 0.69693f*tonal->mem[i+8] - 1.4349f*tonal->cmean[i];
for (i=0;i<4;i++)
tonal->cmean[i] = (1-alpha)*tonal->cmean[i] + alpha*BFCC[i];
for (i=0;i<4;i++)
features[4+i] = 0.63246f*(BFCC[i]-tonal->mem[i+24]) + 0.31623f*(tonal->mem[i]-tonal->mem[i+16]);
for (i=0;i<3;i++)
features[8+i] = 0.53452f*(BFCC[i]+tonal->mem[i+24]) - 0.26726f*(tonal->mem[i]+tonal->mem[i+16]) -0.53452f*tonal->mem[i+8];
if (tonal->count > 5)
{
for (i=0;i<9;i++)
tonal->std[i] = (1-alpha)*tonal->std[i] + alpha*features[i]*features[i];
}
for (i=0;i<4;i++)
features[i] = BFCC[i]-midE[i];
for (i=0;i<8;i++)
{
tonal->mem[i+24] = tonal->mem[i+16];
tonal->mem[i+16] = tonal->mem[i+8];
tonal->mem[i+8] = tonal->mem[i];
tonal->mem[i] = BFCC[i];
}
for (i=0;i<9;i++)
features[11+i] = (float)sqrt(tonal->std[i]) - std_feature_bias[i];
features[18] = spec_variability - 0.78f;
features[20] = info->tonality - 0.154723f;
features[21] = info->activity - 0.724643f;
features[22] = frame_stationarity - 0.743717f;
features[23] = info->tonality_slope + 0.069216f;
features[24] = tonal->lowECount - 0.067930f;
mlp_process(&net, features, frame_probs);
frame_probs[0] = .5f*(frame_probs[0]+1);
/* Curve fitting between the MLP probability and the actual probability */
/*frame_probs[0] = .01f + 1.21f*frame_probs[0]*frame_probs[0] - .23f*(float)pow(frame_probs[0], 10);*/
/* Probability of active audio (as opposed to silence) */
frame_probs[1] = .5f*frame_probs[1]+.5f;
frame_probs[1] *= frame_probs[1];
/* Probability of speech or music vs noise */
info->activity_probability = frame_probs[1];
/*printf("%f %f\n", frame_probs[0], frame_probs[1]);*/
{
/* Probability of state transition */
float tau;
/* Represents independence of the MLP probabilities, where
beta=1 means fully independent. */
float beta;
/* Denormalized probability of speech (p0) and music (p1) after update */
float p0, p1;
/* Probabilities for "all speech" and "all music" */
float s0, m0;
/* Probability sum for renormalisation */
float psum;
/* Instantaneous probability of speech and music, with beta pre-applied. */
float speech0;
float music0;
float p, q;
/* More silence transitions for speech than for music. */
tau = .001f*tonal->music_prob + .01f*(1-tonal->music_prob);
p = MAX16(.05f,MIN16(.95f,frame_probs[1]));
q = MAX16(.05f,MIN16(.95f,tonal->vad_prob));
beta = .02f+.05f*ABS16(p-q)/(p*(1-q)+q*(1-p));
/* p0 and p1 are the probabilities of speech and music at this frame
using only information from previous frame and applying the
state transition model */
p0 = (1-tonal->vad_prob)*(1-tau) + tonal->vad_prob *tau;
p1 = tonal->vad_prob *(1-tau) + (1-tonal->vad_prob)*tau;
/* We apply the current probability with exponent beta to work around
the fact that the probability estimates aren't independent. */
p0 *= (float)pow(1-frame_probs[1], beta);
p1 *= (float)pow(frame_probs[1], beta);
/* Normalise the probabilities to get the Marokv probability of music. */
tonal->vad_prob = p1/(p0+p1);
info->vad_prob = tonal->vad_prob;
/* Consider that silence has a 50-50 probability of being speech or music. */
frame_probs[0] = tonal->vad_prob*frame_probs[0] + (1-tonal->vad_prob)*.5f;
/* One transition every 3 minutes of active audio */
tau = .0001f;
/* Adapt beta based on how "unexpected" the new prob is */
p = MAX16(.05f,MIN16(.95f,frame_probs[0]));
q = MAX16(.05f,MIN16(.95f,tonal->music_prob));
beta = .02f+.05f*ABS16(p-q)/(p*(1-q)+q*(1-p));
/* p0 and p1 are the probabilities of speech and music at this frame
using only information from previous frame and applying the
state transition model */
p0 = (1-tonal->music_prob)*(1-tau) + tonal->music_prob *tau;
p1 = tonal->music_prob *(1-tau) + (1-tonal->music_prob)*tau;
/* We apply the current probability with exponent beta to work around
the fact that the probability estimates aren't independent. */
p0 *= (float)pow(1-frame_probs[0], beta);
p1 *= (float)pow(frame_probs[0], beta);
/* Normalise the probabilities to get the Marokv probability of music. */
tonal->music_prob = p1/(p0+p1);
info->music_prob = tonal->music_prob;
/*printf("%f %f %f %f\n", frame_probs[0], frame_probs[1], tonal->music_prob, tonal->vad_prob);*/
/* This chunk of code deals with delayed decision. */
psum=1e-20f;
/* Instantaneous probability of speech and music, with beta pre-applied. */
speech0 = (float)pow(1-frame_probs[0], beta);
music0 = (float)pow(frame_probs[0], beta);
if (tonal->count==1)
{
if (tonal->application == OPUS_APPLICATION_VOIP)
tonal->pmusic[0] = .1f;
else
tonal->pmusic[0] = .625f;
tonal->pspeech[0] = 1-tonal->pmusic[0];
}
/* Updated probability of having only speech (s0) or only music (m0),
before considering the new observation. */
s0 = tonal->pspeech[0] + tonal->pspeech[1];
m0 = tonal->pmusic [0] + tonal->pmusic [1];
/* Updates s0 and m0 with instantaneous probability. */
tonal->pspeech[0] = s0*(1-tau)*speech0;
tonal->pmusic [0] = m0*(1-tau)*music0;
/* Propagate the transition probabilities */
for (i=1;i<DETECT_SIZE-1;i++)
{
tonal->pspeech[i] = tonal->pspeech[i+1]*speech0;
tonal->pmusic [i] = tonal->pmusic [i+1]*music0;
}
/* Probability that the latest frame is speech, when all the previous ones were music. */
tonal->pspeech[DETECT_SIZE-1] = m0*tau*speech0;
/* Probability that the latest frame is music, when all the previous ones were speech. */
tonal->pmusic [DETECT_SIZE-1] = s0*tau*music0;
/* Renormalise probabilities to 1 */
for (i=0;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i] + tonal->pmusic[i];
psum = 1.f/psum;
for (i=0;i<DETECT_SIZE;i++)
{
tonal->pspeech[i] *= psum;
tonal->pmusic [i] *= psum;
}
psum = tonal->pmusic[0];
for (i=1;i<DETECT_SIZE;i++)
psum += tonal->pspeech[i];
/* Estimate our confidence in the speech/music decisions */
if (frame_probs[1]>.75)
{
if (tonal->music_prob>.9)
{
float adapt;
adapt = 1.f/(++tonal->music_confidence_count);
tonal->music_confidence_count = IMIN(tonal->music_confidence_count, 500);
tonal->music_confidence += adapt*MAX16(-.2f,frame_probs[0]-tonal->music_confidence);
}
if (tonal->music_prob<.1)
{
float adapt;
adapt = 1.f/(++tonal->speech_confidence_count);
tonal->speech_confidence_count = IMIN(tonal->speech_confidence_count, 500);
tonal->speech_confidence += adapt*MIN16(.2f,frame_probs[0]-tonal->speech_confidence);
}
}
}
tonal->last_music = tonal->music_prob>.5f;
#ifdef MLP_TRAINING
for (i=0;i<25;i++)
printf("%f ", features[i]);
printf("\n");
#endif
info->bandwidth = bandwidth;
tonal->prev_bandwidth = bandwidth;
/*printf("%d %d\n", info->bandwidth, info->opus_bandwidth);*/
info->noisiness = frame_noisiness;
info->valid = 1;
RESTORE_STACK;
}