int speex_preprocess(SpeexPreprocessState *st, short *x, float *echo)
{
int i;
int is_speech=1;
float mean_post=0;
float mean_prior=0;
int N = st->ps_size;
int N3 = 2*N - st->frame_size;
int N4 = st->frame_size - N3;
float scale=.5/N;
float *ps=st->ps;
float Zframe=0, Pframe;
preprocess_analysis(st, x);
update_noise_prob(st);
st->nb_preprocess++;
/* Noise estimation always updated for the 20 first times */
if (st->nb_adapt<10)
{
update_noise(st, ps, echo);
}
/* Deal with residual echo if provided */
if (echo)
for (i=1;i<N;i++)
st->echo_noise[i] = (.7*st->echo_noise[i] + .3* echo[i]);
/* Compute a posteriori SNR */
for (i=1;i<N;i++)
{
st->post[i] = ps[i]/(1+st->noise[i]+st->echo_noise[i]) - 1;
if (st->post[i]>100)
st->post[i]=100;
/*if (st->post[i]<0)
st->post[i]=0;*/
mean_post+=st->post[i];
}
mean_post /= N;
if (mean_post<0)
mean_post=0;
/* Special case for first frame */
if (st->nb_adapt==1)
for (i=1;i<N;i++)
st->old_ps[i] = ps[i];
/* Compute a priori SNR */
{
/* A priori update rate */
float gamma;
float min_gamma=0.12;
gamma = 1.0/st->nb_preprocess;
/*Make update rate smaller when there's no speech*/
#if 0
if (mean_post<3.5 && mean_prior < 1)
min_gamma *= (mean_post+.5);
else
min_gamma *= 4.;
#else
min_gamma = .1*fabs(mean_prior - mean_post)*fabs(mean_prior - mean_post);
if (min_gamma>.15)
min_gamma = .15;
if (min_gamma<.02)
min_gamma = .02;
#endif
/*min_gamma = .08;*/
/*if (gamma<min_gamma)*/
gamma=min_gamma;
for (i=1;i<N;i++)
{
/* A priori SNR update */
st->prior[i] = gamma*max(0.0,st->post[i]) +
(1-gamma)*st->gain[i]*st->gain[i]*st->old_ps[i]/(1+st->noise[i]+st->echo_noise[i]);
if (st->prior[i]>100)
st->prior[i]=100;
mean_prior+=st->prior[i];
}
}
mean_prior /= N;
#if 0
for (i=0;i<N;i++)
{
fprintf (stderr, "%f ", st->prior[i]);
}
fprintf (stderr, "\n");
#endif
/*fprintf (stderr, "%f %f\n", mean_prior,mean_post);*/
if (st->nb_preprocess>=20)
{
int do_update = 0;
float noise_ener=0, sig_ener=0;
/* If SNR is low (both a priori and a posteriori), update the noise estimate*/
/*if (mean_prior<.23 && mean_post < .5)*/
if (mean_prior<.23 && mean_post < .5)
do_update = 1;
for (i=1;i<N;i++)
{
noise_ener += st->noise[i];
sig_ener += ps[i];
}
if (noise_ener > 3*sig_ener)
do_update = 1;
/*do_update = 0;*/
if (do_update)
{
st->consec_noise++;
} else {
st->consec_noise=0;
}
}
if (st->vad_enabled)
is_speech = speex_compute_vad(st, ps, mean_prior, mean_post);
if (st->consec_noise>=3)
{
update_noise(st, st->old_ps, echo);
} else {
for (i=1;i<N-1;i++)
{
if (st->update_prob[i]<.5)
st->noise[i] = .90*st->noise[i] + .1*st->ps[i];
}
}
for (i=1;i<N;i++)
{
st->zeta[i] = .7*st->zeta[i] + .3*st->prior[i];
}
{
int freq_start = (int)(300.0*2*N/st->sampling_rate);
int freq_end = (int)(2000.0*2*N/st->sampling_rate);
for (i=freq_start;i<freq_end;i++)
{
Zframe += st->zeta[i];
}
}
Zframe /= N;
if (Zframe<ZMIN)
{
Pframe = 0;
} else {
if (Zframe > 1.5*st->Zlast)
{
Pframe = 1;
st->Zpeak = Zframe;
if (st->Zpeak > 10)
st->Zpeak = 10;
if (st->Zpeak < 1)
st->Zpeak = 1;
} else {
if (Zframe < st->Zpeak*ZMIN)
{
Pframe = 0;
} else if (Zframe > st->Zpeak*ZMAX)
{
Pframe = 1;
} else {
Pframe = log(Zframe/(st->Zpeak*ZMIN)) / log(ZMAX/ZMIN);
}
}
}
st->Zlast = Zframe;
/*fprintf (stderr, "%f\n", Pframe);*/
/* Compute gain according to the Ephraim-Malah algorithm */
ephraim_malah(st,N,Pframe);
if (st->agc_enabled)
speex_compute_agc(st, mean_prior);
#if 0
if (!is_speech)
{
for (i=0;i<N;i++)
st->gain2[i] = 0;
}
#if 0
else {
for (i=0;i<N;i++)
st->gain2[i] = 1;
}
#endif
#endif
/* PERF: 14% when only vad is enabled [7.0 vs 8.2 sec] */
if(st->agc_enabled || st->denoise_enabled) {
/* Apply computed gain */
for (i=1;i<N;i++)
{
st->frame[2*i-1] *= st->gain2[i];
st->frame[2*i] *= st->gain2[i];
}
/* Get rid of the DC and very low frequencies */
st->frame[0]=0;
st->frame[1]=0;
st->frame[2]=0;
/* Nyquist frequency is mostly useless too */
st->frame[2*N-1]=0;
/* Inverse FFT with 1/N scaling */
drft_backward(st->fft_lookup, st->frame);
for (i=0;i<2*N;i++)
st->frame[i] *= scale;
{
float max_sample=0;
for (i=0;i<2*N;i++)
if (fabs(st->frame[i])>max_sample)
max_sample = fabs(st->frame[i]);
if (max_sample>28000)
{
float damp = 28000./max_sample;
for (i=0;i<2*N;i++)
st->frame[i] *= damp;
}
}
for (i=0;i<2*N;i++)
st->frame[i] *= st->window[i];
/* Perform overlap and add */
for (i=0;i<N3;i++)
x[i] = st->outbuf[i] + st->frame[i];
for (i=0;i<N4;i++)
x[N3+i] = st->frame[N3+i];
/* Update outbuf */
for (i=0;i<N3;i++)
st->outbuf[i] = st->frame[st->frame_size+i];
}
/* Save old power spectrum */
for (i=1;i<N;i++)
st->old_ps[i] = ps[i];
return is_speech;
}
int speex_preprocess(SpeexPreprocessState *st, spx_int16_t *x, float *echo)
{
int i;
int is_speech=1;
float mean_post=0;
float mean_prior=0;
int N = st->ps_size;
int N3 = 2*N - st->frame_size;
int N4 = st->frame_size - N3;
float scale=.5f/N;
float *ps=st->ps;
float Zframe=0, Pframe;
preprocess_analysis(st, x);
update_noise_prob(st);
st->nb_preprocess++;
/* Noise estimation always updated for the 20 first times */
if (st->nb_adapt<10)
{
update_noise(st, ps, echo);
}
/* Deal with residual echo if provided */
if (echo)
for (i=1;i<N;i++)
st->echo_noise[i] = (.3f*st->echo_noise[i] + echo[i]);
/* Compute a posteriori SNR */
for (i=1;i<N;i++)
{
st->post[i] = ps[i]/(1.f+NOISE_OVERCOMPENS*st->noise[i]+st->echo_noise[i]+st->reverb_estimate[i]) - 1.f;
if (st->post[i]>100.f)
st->post[i]=100.f;
/*if (st->post[i]<0)
st->post[i]=0;*/
mean_post+=st->post[i];
}
mean_post /= N;
if (mean_post<0.f)
mean_post=0.f;
/* Special case for first frame */
if (st->nb_adapt==1)
for (i=1;i<N;i++)
st->old_ps[i] = ps[i];
/* Compute a priori SNR */
{
/* A priori update rate */
float gamma;
float min_gamma=0.12f;
gamma = 1.0f/st->nb_preprocess;
/*Make update rate smaller when there's no speech*/
#if 0
if (mean_post<3.5 && mean_prior < 1)
min_gamma *= (mean_post+.5);
else
min_gamma *= 4.;
#else
min_gamma = .1f*fabs(mean_prior - mean_post)*fabs(mean_prior - mean_post);
if (min_gamma>.15f)
min_gamma = .15f;
if (min_gamma<.02f)
min_gamma = .02f;
#endif
/*min_gamma = .08;*/
/*if (gamma<min_gamma)*/
gamma=min_gamma;
gamma = .1;
for (i=1;i<N;i++)
{
/* A priori SNR update */
st->prior[i] = gamma*max(0.0f,st->post[i]) +
(1.f-gamma)*st->gain[i]*st->gain[i]*st->old_ps[i]/(1.f+NOISE_OVERCOMPENS*st->noise[i]+st->echo_noise[i]+st->reverb_estimate[i]);
if (st->prior[i]>100.f)
st->prior[i]=100.f;
mean_prior+=st->prior[i];
}
}
mean_prior /= N;
#if 0
for (i=0;i<N;i++)
{
fprintf (stderr, "%f ", st->prior[i]);
}
fprintf (stderr, "\n");
#endif
/*fprintf (stderr, "%f %f\n", mean_prior,mean_post);*/
if (st->nb_preprocess>=20)
{
int do_update = 0;
float noise_ener=0, sig_ener=0;
/* If SNR is low (both a priori and a posteriori), update the noise estimate*/
/*if (mean_prior<.23 && mean_post < .5)*/
if (mean_prior<.23f && mean_post < .5f)
do_update = 1;
for (i=1;i<N;i++)
{
noise_ener += st->noise[i];
sig_ener += ps[i];
}
if (noise_ener > 3.f*sig_ener)
do_update = 1;
/*do_update = 0;*/
if (do_update)
{
st->consec_noise++;
} else {
st->consec_noise=0;
}
}
if (st->vad_enabled)
is_speech = speex_compute_vad(st, ps, mean_prior, mean_post);
if (st->consec_noise>=3)
{
update_noise(st, st->old_ps, echo);
} else {
for (i=1;i<N-1;i++)
{
if (st->update_prob[i]<.5f || st->ps[i] < st->noise[i])
{
if (echo)
st->noise[i] = .90f*st->noise[i] + .1f*max(1.0f,st->ps[i]-echo[i]);
else
st->noise[i] = .90f*st->noise[i] + .1f*st->ps[i];
}
}
}
for (i=1;i<N;i++)
{
st->zeta[i] = .7f*st->zeta[i] + .3f*st->prior[i];
}
{
int freq_start = (int)(300.0f*2.f*N/st->sampling_rate);
int freq_end = (int)(2000.0f*2.f*N/st->sampling_rate);
for (i=freq_start;i<freq_end;i++)
{
Zframe += st->zeta[i];
}
}
Zframe /= N;
if (Zframe<ZMIN)
{
Pframe = 0;
} else {
if (Zframe > 1.5f*st->Zlast)
{
Pframe = 1.f;
st->Zpeak = Zframe;
if (st->Zpeak > 10.f)
st->Zpeak = 10.f;
if (st->Zpeak < 1.f)
st->Zpeak = 1.f;
} else {
if (Zframe < st->Zpeak*ZMIN)
{
Pframe = 0;
} else if (Zframe > st->Zpeak*ZMAX)
{
Pframe = 1;
} else {
Pframe = log(Zframe/(st->Zpeak*ZMIN)) / log(ZMAX/ZMIN);
}
}
}
st->Zlast = Zframe;
/*fprintf (stderr, "%f\n", Pframe);*/
/* Compute gain according to the Ephraim-Malah algorithm */
for (i=1;i<N;i++)
{
float MM;
float theta;
float prior_ratio;
float p, q;
float zeta1;
float P1;
prior_ratio = st->prior[i]/(1.0001f+st->prior[i]);
theta = (1.f+st->post[i])*prior_ratio;
if (i==1 || i==N-1)
zeta1 = st->zeta[i];
else
zeta1 = .25f*st->zeta[i-1] + .5f*st->zeta[i] + .25f*st->zeta[i+1];
if (zeta1<ZMIN)
P1 = 0.f;
else if (zeta1>ZMAX)
P1 = 1.f;
else
P1 = LOG_MIN_MAX_1 * log(ZMIN_1*zeta1);
/*P1 = log(zeta1/ZMIN)/log(ZMAX/ZMIN);*/
/* FIXME: add global prob (P2) */
q = 1-Pframe*P1;
q = 1-P1;
if (q>.95f)
q=.95f;
p=1.f/(1.f + (q/(1.f-q))*(1.f+st->prior[i])*exp(-theta));
/*p=1;*/
#if 0
/* log-spectral magnitude estimator */
if (theta<6)
MM = 0.74082*pow(theta+1,.61)/sqrt(.0001+theta);
else
MM=1;
#else
/* Optimal estimator for loudness domain */
MM = hypergeom_gain(theta);
#endif
st->gain[i] = prior_ratio * MM;
/*Put some (very arbitraty) limit on the gain*/
if (st->gain[i]>2.f)
{
st->gain[i]=2.f;
}
st->reverb_estimate[i] = st->reverb_decay*st->reverb_estimate[i] + st->reverb_decay*st->reverb_level*st->gain[i]*st->gain[i]*st->ps[i];
if (st->denoise_enabled)
{
st->gain2[i]=p*p*st->gain[i];
} else {
st->gain2[i]=1.f;
}
}
st->gain2[0]=st->gain[0]=0.f;
st->gain2[N-1]=st->gain[N-1]=0.f;
if (st->agc_enabled)
speex_compute_agc(st, mean_prior);
#if 0
if (!is_speech)
{
for (i=0;i<N;i++)
st->gain2[i] = 0;
}
#if 0
else {
for (i=0;i<N;i++)
st->gain2[i] = 1;
}
#endif
#endif
/* Apply computed gain */
for (i=1;i<N;i++)
{
st->frame[2*i-1] *= st->gain2[i];
st->frame[2*i] *= st->gain2[i];
}
/* Get rid of the DC and very low frequencies */
st->frame[0]=0;
st->frame[1]=0;
st->frame[2]=0;
/* Nyquist frequency is mostly useless too */
st->frame[2*N-1]=0;
/* Inverse FFT with 1/N scaling */
spx_drft_backward(st->fft_lookup, st->frame);
for (i=0;i<2*N;i++)
st->frame[i] *= scale;
{
float max_sample=0;
for (i=0;i<2*N;i++)
if (fabs(st->frame[i])>max_sample)
max_sample = fabs(st->frame[i]);
if (max_sample>28000.f)
{
float damp = 28000.f/max_sample;
for (i=0;i<2*N;i++)
st->frame[i] *= damp;
}
}
for (i=0;i<2*N;i++)
st->frame[i] *= st->window[i];
/* Perform overlap and add */
for (i=0;i<N3;i++)
x[i] = st->outbuf[i] + st->frame[i];
for (i=0;i<N4;i++)
x[N3+i] = st->frame[N3+i];
/* Update outbuf */
for (i=0;i<N3;i++)
st->outbuf[i] = st->frame[st->frame_size+i];
/* Save old power spectrum */
for (i=1;i<N;i++)
st->old_ps[i] = ps[i];
return is_speech;
}
int speex_preprocess(SpeexPreprocessState *st, spx_int16_t *x, spx_int32_t *echo)
{
int i;
int is_speech=1;
float mean_post=0;
float mean_prior=0;
int N = st->ps_size;
int N3 = 2*N - st->frame_size;
int N4 = st->frame_size - N3;
float scale=.5f/N;
float *ps=st->ps;
float Zframe=0, Pframe;
preprocess_analysis(st, x);
update_noise_prob(st);
st->nb_preprocess++;
/* Noise estimation always updated for the 20 first times */
if (st->nb_adapt<10)
{
update_noise(st, ps, echo);
}
/* Deal with residual echo if provided */
if (echo)
for (i=1;i<N;i++)
st->echo_noise[i] = (.3f*st->echo_noise[i] + st->frame_size*st->frame_size*1.0*echo[i]);
/* Compute a posteriori SNR */
for (i=1;i<N;i++)
{
float tot_noise = 1.f+ NOISE_OVERCOMPENS*st->noise[i] + st->echo_noise[i] + st->reverb_estimate[i];
st->post[i] = ps[i]/tot_noise - 1.f;
if (st->post[i]>100.f)
st->post[i]=100.f;
/*if (st->post[i]<0)
st->post[i]=0;*/
mean_post+=st->post[i];
}
mean_post /= N;
if (mean_post<0.f)
mean_post=0.f;
/* Special case for first frame */
if (st->nb_adapt==1)
for (i=1;i<N;i++)
st->old_ps[i] = ps[i];
/* Compute a priori SNR */
{
/* A priori update rate */
for (i=1;i<N;i++)
{
float gamma = .15+.85*st->prior[i]*st->prior[i]/((1+st->prior[i])*(1+st->prior[i]));
float tot_noise = 1.f+ NOISE_OVERCOMPENS*st->noise[i] + st->echo_noise[i] + st->reverb_estimate[i];
/* A priori SNR update */
st->prior[i] = gamma*max(0.0f,st->post[i]) +
(1.f-gamma)* (.8*st->gain[i]*st->gain[i]*st->old_ps[i]/tot_noise + .2*st->prior[i]);
if (st->prior[i]>100.f)
st->prior[i]=100.f;
mean_prior+=st->prior[i];
}
}
mean_prior /= N;
#if 0
for (i=0;i<N;i++)
{
fprintf (stderr, "%f ", st->prior[i]);
}
fprintf (stderr, "\n");
#endif
/*fprintf (stderr, "%f %f\n", mean_prior,mean_post);*/
if (st->nb_preprocess>=20)
{
int do_update = 0;
float noise_ener=0, sig_ener=0;
/* If SNR is low (both a priori and a posteriori), update the noise estimate*/
/*if (mean_prior<.23 && mean_post < .5)*/
if (mean_prior<.23f && mean_post < .5f)
do_update = 1;
for (i=1;i<N;i++)
{
noise_ener += st->noise[i];
sig_ener += ps[i];
}
if (noise_ener > 3.f*sig_ener)
do_update = 1;
/*do_update = 0;*/
if (do_update)
{
st->consec_noise++;
} else {
st->consec_noise=0;
}
}
if (st->vad_enabled)
is_speech = speex_compute_vad(st, ps, mean_prior, mean_post);
if (st->consec_noise>=3)
{
update_noise(st, st->old_ps, echo);
} else {
for (i=1;i<N-1;i++)
{
if (st->update_prob[i]<.5f/* || st->ps[i] < st->noise[i]*/)
{
if (echo)
st->noise[i] = .95f*st->noise[i] + .05f*max(1.0f,st->ps[i]-st->frame_size*st->frame_size*1.0*echo[i]);
else
st->noise[i] = .95f*st->noise[i] + .05f*st->ps[i];
}
}
}
for (i=1;i<N;i++)
{
st->zeta[i] = .7f*st->zeta[i] + .3f*st->prior[i];
}
{
int freq_start = (int)(300.0f*2.f*N/st->sampling_rate);
int freq_end = (int)(2000.0f*2.f*N/st->sampling_rate);
for (i=freq_start;i<freq_end;i++)
{
Zframe += st->zeta[i];
}
Zframe /= (freq_end-freq_start);
}
st->Zlast = Zframe;
Pframe = qcurve(Zframe);
/*fprintf (stderr, "%f\n", Pframe);*/
/* Compute gain according to the Ephraim-Malah algorithm */
for (i=1;i<N;i++)
{
float MM;
float theta;
float prior_ratio;
float p, q;
float zeta1;
float P1;
prior_ratio = st->prior[i]/(1.0001f+st->prior[i]);
theta = (1.f+st->post[i])*prior_ratio;
if (i==1 || i==N-1)
zeta1 = st->zeta[i];
else
zeta1 = .25f*st->zeta[i-1] + .5f*st->zeta[i] + .25f*st->zeta[i+1];
P1 = qcurve (zeta1);
/* FIXME: add global prob (P2) */
q = 1-Pframe*P1;
q = 1-P1;
if (q>.95f)
q=.95f;
p=1.f/(1.f + (q/(1.f-q))*(1.f+st->prior[i])*exp(-theta));
/*p=1;*/
/* Optimal estimator for loudness domain */
MM = hypergeom_gain(theta);
st->gain[i] = prior_ratio * MM;
/*Put some (very arbitraty) limit on the gain*/
if (st->gain[i]>2.f)
{
st->gain[i]=2.f;
}
st->reverb_estimate[i] = st->reverb_decay*st->reverb_estimate[i] + st->reverb_decay*st->reverb_level*st->gain[i]*st->gain[i]*st->ps[i];
if (st->denoise_enabled)
{
/*st->gain2[i] = p*p*st->gain[i];*/
st->gain2[i]=(p*sqrt(st->gain[i])+.2*(1-p)) * (p*sqrt(st->gain[i])+.2*(1-p));
/*st->gain2[i] = pow(st->gain[i], p) * pow(.1f,1.f-p);*/
} else {
st->gain2[i]=1.f;
}
}
st->gain2[0]=st->gain[0]=0.f;
st->gain2[N-1]=st->gain[N-1]=0.f;
/*
for (i=30;i<N-2;i++)
{
st->gain[i] = st->gain2[i]*st->gain2[i] + (1-st->gain2[i])*.333*(.6*st->gain2[i-1]+st->gain2[i]+.6*st->gain2[i+1]+.4*st->gain2[i-2]+.4*st->gain2[i+2]);
}
for (i=30;i<N-2;i++)
st->gain2[i] = st->gain[i];
*/
if (st->agc_enabled)
speex_compute_agc(st, mean_prior);
#if 0
if (!is_speech)
{
for (i=0;i<N;i++)
st->gain2[i] = 0;
}
#if 0
else {
for (i=0;i<N;i++)
st->gain2[i] = 1;
}
#endif
#endif
/* Apply computed gain */
for (i=1;i<N;i++)
{
st->frame[2*i-1] *= st->gain2[i];
st->frame[2*i] *= st->gain2[i];
}
/* Get rid of the DC and very low frequencies */
st->frame[0]=0;
st->frame[1]=0;
st->frame[2]=0;
/* Nyquist frequency is mostly useless too */
st->frame[2*N-1]=0;
/* Inverse FFT with 1/N scaling */
spx_drft_backward(st->fft_lookup, st->frame);
for (i=0;i<2*N;i++)
st->frame[i] *= scale;
{
float max_sample=0;
for (i=0;i<2*N;i++)
if (fabs(st->frame[i])>max_sample)
max_sample = fabs(st->frame[i]);
if (max_sample>28000.f)
{
float damp = 28000.f/max_sample;
for (i=0;i<2*N;i++)
st->frame[i] *= damp;
}
}
for (i=0;i<2*N;i++)
st->frame[i] *= st->window[i];
/* Perform overlap and add */
for (i=0;i<N3;i++)
x[i] = st->outbuf[i] + st->frame[i];
for (i=0;i<N4;i++)
x[N3+i] = st->frame[N3+i];
/* Update outbuf */
for (i=0;i<N3;i++)
st->outbuf[i] = st->frame[st->frame_size+i];
/* Save old power spectrum */
for (i=1;i<N;i++)
st->old_ps[i] = ps[i];
return is_speech;
}