michael@0: /* Copyright (c) 2011 Xiph.Org Foundation
michael@0:    Written by Jean-Marc Valin */
michael@0: /*
michael@0:    Redistribution and use in source and binary forms, with or without
michael@0:    modification, are permitted provided that the following conditions
michael@0:    are met:
michael@0: 
michael@0:    - Redistributions of source code must retain the above copyright
michael@0:    notice, this list of conditions and the following disclaimer.
michael@0: 
michael@0:    - Redistributions in binary form must reproduce the above copyright
michael@0:    notice, this list of conditions and the following disclaimer in the
michael@0:    documentation and/or other materials provided with the distribution.
michael@0: 
michael@0:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
michael@0:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
michael@0:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
michael@0:    A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
michael@0:    CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
michael@0:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
michael@0:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
michael@0:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
michael@0:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
michael@0:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
michael@0:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
michael@0: */
michael@0: 
michael@0: #ifdef HAVE_CONFIG_H
michael@0: #include "config.h"
michael@0: #endif
michael@0: 
michael@0: #include "kiss_fft.h"
michael@0: #include "celt.h"
michael@0: #include "modes.h"
michael@0: #include "arch.h"
michael@0: #include "quant_bands.h"
michael@0: #include <stdio.h>
michael@0: #include "analysis.h"
michael@0: #include "mlp.h"
michael@0: #include "stack_alloc.h"
michael@0: 
michael@0: extern const MLP net;
michael@0: 
michael@0: #ifndef M_PI
michael@0: #define M_PI 3.141592653
michael@0: #endif
michael@0: 
michael@0: static const float dct_table[128] = {
michael@0:         0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f,
michael@0:         0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f,
michael@0:         0.351851f, 0.338330f, 0.311806f, 0.273300f, 0.224292f, 0.166664f, 0.102631f, 0.034654f,
michael@0:        -0.034654f,-0.102631f,-0.166664f,-0.224292f,-0.273300f,-0.311806f,-0.338330f,-0.351851f,
michael@0:         0.346760f, 0.293969f, 0.196424f, 0.068975f,-0.068975f,-0.196424f,-0.293969f,-0.346760f,
michael@0:        -0.346760f,-0.293969f,-0.196424f,-0.068975f, 0.068975f, 0.196424f, 0.293969f, 0.346760f,
michael@0:         0.338330f, 0.224292f, 0.034654f,-0.166664f,-0.311806f,-0.351851f,-0.273300f,-0.102631f,
michael@0:         0.102631f, 0.273300f, 0.351851f, 0.311806f, 0.166664f,-0.034654f,-0.224292f,-0.338330f,
michael@0:         0.326641f, 0.135299f,-0.135299f,-0.326641f,-0.326641f,-0.135299f, 0.135299f, 0.326641f,
michael@0:         0.326641f, 0.135299f,-0.135299f,-0.326641f,-0.326641f,-0.135299f, 0.135299f, 0.326641f,
michael@0:         0.311806f, 0.034654f,-0.273300f,-0.338330f,-0.102631f, 0.224292f, 0.351851f, 0.166664f,
michael@0:        -0.166664f,-0.351851f,-0.224292f, 0.102631f, 0.338330f, 0.273300f,-0.034654f,-0.311806f,
michael@0:         0.293969f,-0.068975f,-0.346760f,-0.196424f, 0.196424f, 0.346760f, 0.068975f,-0.293969f,
michael@0:        -0.293969f, 0.068975f, 0.346760f, 0.196424f,-0.196424f,-0.346760f,-0.068975f, 0.293969f,
michael@0:         0.273300f,-0.166664f,-0.338330f, 0.034654f, 0.351851f, 0.102631f,-0.311806f,-0.224292f,
michael@0:         0.224292f, 0.311806f,-0.102631f,-0.351851f,-0.034654f, 0.338330f, 0.166664f,-0.273300f,
michael@0: };
michael@0: 
michael@0: static const float analysis_window[240] = {
michael@0:       0.000043f, 0.000171f, 0.000385f, 0.000685f, 0.001071f, 0.001541f, 0.002098f, 0.002739f,
michael@0:       0.003466f, 0.004278f, 0.005174f, 0.006156f, 0.007222f, 0.008373f, 0.009607f, 0.010926f,
michael@0:       0.012329f, 0.013815f, 0.015385f, 0.017037f, 0.018772f, 0.020590f, 0.022490f, 0.024472f,
michael@0:       0.026535f, 0.028679f, 0.030904f, 0.033210f, 0.035595f, 0.038060f, 0.040604f, 0.043227f,
michael@0:       0.045928f, 0.048707f, 0.051564f, 0.054497f, 0.057506f, 0.060591f, 0.063752f, 0.066987f,
michael@0:       0.070297f, 0.073680f, 0.077136f, 0.080665f, 0.084265f, 0.087937f, 0.091679f, 0.095492f,
michael@0:       0.099373f, 0.103323f, 0.107342f, 0.111427f, 0.115579f, 0.119797f, 0.124080f, 0.128428f,
michael@0:       0.132839f, 0.137313f, 0.141849f, 0.146447f, 0.151105f, 0.155823f, 0.160600f, 0.165435f,
michael@0:       0.170327f, 0.175276f, 0.180280f, 0.185340f, 0.190453f, 0.195619f, 0.200838f, 0.206107f,
michael@0:       0.211427f, 0.216797f, 0.222215f, 0.227680f, 0.233193f, 0.238751f, 0.244353f, 0.250000f,
michael@0:       0.255689f, 0.261421f, 0.267193f, 0.273005f, 0.278856f, 0.284744f, 0.290670f, 0.296632f,
michael@0:       0.302628f, 0.308658f, 0.314721f, 0.320816f, 0.326941f, 0.333097f, 0.339280f, 0.345492f,
michael@0:       0.351729f, 0.357992f, 0.364280f, 0.370590f, 0.376923f, 0.383277f, 0.389651f, 0.396044f,
michael@0:       0.402455f, 0.408882f, 0.415325f, 0.421783f, 0.428254f, 0.434737f, 0.441231f, 0.447736f,
michael@0:       0.454249f, 0.460770f, 0.467298f, 0.473832f, 0.480370f, 0.486912f, 0.493455f, 0.500000f,
michael@0:       0.506545f, 0.513088f, 0.519630f, 0.526168f, 0.532702f, 0.539230f, 0.545751f, 0.552264f,
michael@0:       0.558769f, 0.565263f, 0.571746f, 0.578217f, 0.584675f, 0.591118f, 0.597545f, 0.603956f,
michael@0:       0.610349f, 0.616723f, 0.623077f, 0.629410f, 0.635720f, 0.642008f, 0.648271f, 0.654508f,
michael@0:       0.660720f, 0.666903f, 0.673059f, 0.679184f, 0.685279f, 0.691342f, 0.697372f, 0.703368f,
michael@0:       0.709330f, 0.715256f, 0.721144f, 0.726995f, 0.732807f, 0.738579f, 0.744311f, 0.750000f,
michael@0:       0.755647f, 0.761249f, 0.766807f, 0.772320f, 0.777785f, 0.783203f, 0.788573f, 0.793893f,
michael@0:       0.799162f, 0.804381f, 0.809547f, 0.814660f, 0.819720f, 0.824724f, 0.829673f, 0.834565f,
michael@0:       0.839400f, 0.844177f, 0.848895f, 0.853553f, 0.858151f, 0.862687f, 0.867161f, 0.871572f,
michael@0:       0.875920f, 0.880203f, 0.884421f, 0.888573f, 0.892658f, 0.896677f, 0.900627f, 0.904508f,
michael@0:       0.908321f, 0.912063f, 0.915735f, 0.919335f, 0.922864f, 0.926320f, 0.929703f, 0.933013f,
michael@0:       0.936248f, 0.939409f, 0.942494f, 0.945503f, 0.948436f, 0.951293f, 0.954072f, 0.956773f,
michael@0:       0.959396f, 0.961940f, 0.964405f, 0.966790f, 0.969096f, 0.971321f, 0.973465f, 0.975528f,
michael@0:       0.977510f, 0.979410f, 0.981228f, 0.982963f, 0.984615f, 0.986185f, 0.987671f, 0.989074f,
michael@0:       0.990393f, 0.991627f, 0.992778f, 0.993844f, 0.994826f, 0.995722f, 0.996534f, 0.997261f,
michael@0:       0.997902f, 0.998459f, 0.998929f, 0.999315f, 0.999615f, 0.999829f, 0.999957f, 1.000000f,
michael@0: };
michael@0: 
michael@0: static const int tbands[NB_TBANDS+1] = {
michael@0:        2,  4,  6,  8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56, 68, 80, 96, 120
michael@0: };
michael@0: 
michael@0: static const int extra_bands[NB_TOT_BANDS+1] = {
michael@0:       1, 2,  4,  6,  8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56, 68, 80, 96, 120, 160, 200
michael@0: };
michael@0: 
michael@0: /*static const float tweight[NB_TBANDS+1] = {
michael@0:       .3, .4, .5, .6, .7, .8, .9, 1., 1., 1., 1., 1., 1., 1., .8, .7, .6, .5
michael@0: };*/
michael@0: 
michael@0: #define NB_TONAL_SKIP_BANDS 9
michael@0: 
michael@0: #define cA 0.43157974f
michael@0: #define cB 0.67848403f
michael@0: #define cC 0.08595542f
michael@0: #define cE ((float)M_PI/2)
michael@0: static OPUS_INLINE float fast_atan2f(float y, float x) {
michael@0:    float x2, y2;
michael@0:    /* Should avoid underflow on the values we'll get */
michael@0:    if (ABS16(x)+ABS16(y)<1e-9f)
michael@0:    {
michael@0:       x*=1e12f;
michael@0:       y*=1e12f;
michael@0:    }
michael@0:    x2 = x*x;
michael@0:    y2 = y*y;
michael@0:    if(x2<y2){
michael@0:       float den = (y2 + cB*x2) * (y2 + cC*x2);
michael@0:       if (den!=0)
michael@0:          return -x*y*(y2 + cA*x2) / den + (y<0 ? -cE : cE);
michael@0:       else
michael@0:          return (y<0 ? -cE : cE);
michael@0:    }else{
michael@0:       float den = (x2 + cB*y2) * (x2 + cC*y2);
michael@0:       if (den!=0)
michael@0:          return  x*y*(x2 + cA*y2) / den + (y<0 ? -cE : cE) - (x*y<0 ? -cE : cE);
michael@0:       else
michael@0:          return (y<0 ? -cE : cE) - (x*y<0 ? -cE : cE);
michael@0:    }
michael@0: }
michael@0: 
michael@0: void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int len)
michael@0: {
michael@0:    int pos;
michael@0:    int curr_lookahead;
michael@0:    float psum;
michael@0:    int i;
michael@0: 
michael@0:    pos = tonal->read_pos;
michael@0:    curr_lookahead = tonal->write_pos-tonal->read_pos;
michael@0:    if (curr_lookahead<0)
michael@0:       curr_lookahead += DETECT_SIZE;
michael@0: 
michael@0:    if (len > 480 && pos != tonal->write_pos)
michael@0:    {
michael@0:       pos++;
michael@0:       if (pos==DETECT_SIZE)
michael@0:          pos=0;
michael@0:    }
michael@0:    if (pos == tonal->write_pos)
michael@0:       pos--;
michael@0:    if (pos<0)
michael@0:       pos = DETECT_SIZE-1;
michael@0:    OPUS_COPY(info_out, &tonal->info[pos], 1);
michael@0:    tonal->read_subframe += len/120;
michael@0:    while (tonal->read_subframe>=4)
michael@0:    {
michael@0:       tonal->read_subframe -= 4;
michael@0:       tonal->read_pos++;
michael@0:    }
michael@0:    if (tonal->read_pos>=DETECT_SIZE)
michael@0:       tonal->read_pos-=DETECT_SIZE;
michael@0: 
michael@0:    /* Compensate for the delay in the features themselves.
michael@0:       FIXME: Need a better estimate the 10 I just made up */
michael@0:    curr_lookahead = IMAX(curr_lookahead-10, 0);
michael@0: 
michael@0:    psum=0;
michael@0:    /* Summing the probability of transition patterns that involve music at
michael@0:       time (DETECT_SIZE-curr_lookahead-1) */
michael@0:    for (i=0;i<DETECT_SIZE-curr_lookahead;i++)
michael@0:       psum += tonal->pmusic[i];
michael@0:    for (;i<DETECT_SIZE;i++)
michael@0:       psum += tonal->pspeech[i];
michael@0:    psum = psum*tonal->music_confidence + (1-psum)*tonal->speech_confidence;
michael@0:    /*printf("%f %f %f\n", psum, info_out->music_prob, info_out->tonality);*/
michael@0: 
michael@0:    info_out->music_prob = psum;
michael@0: }
michael@0: 
michael@0: 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)
michael@0: {
michael@0:     int i, b;
michael@0:     const kiss_fft_state *kfft;
michael@0:     VARDECL(kiss_fft_cpx, in);
michael@0:     VARDECL(kiss_fft_cpx, out);
michael@0:     int N = 480, N2=240;
michael@0:     float * OPUS_RESTRICT A = tonal->angle;
michael@0:     float * OPUS_RESTRICT dA = tonal->d_angle;
michael@0:     float * OPUS_RESTRICT d2A = tonal->d2_angle;
michael@0:     VARDECL(float, tonality);
michael@0:     VARDECL(float, noisiness);
michael@0:     float band_tonality[NB_TBANDS];
michael@0:     float logE[NB_TBANDS];
michael@0:     float BFCC[8];
michael@0:     float features[25];
michael@0:     float frame_tonality;
michael@0:     float max_frame_tonality;
michael@0:     /*float tw_sum=0;*/
michael@0:     float frame_noisiness;
michael@0:     const float pi4 = (float)(M_PI*M_PI*M_PI*M_PI);
michael@0:     float slope=0;
michael@0:     float frame_stationarity;
michael@0:     float relativeE;
michael@0:     float frame_probs[2];
michael@0:     float alpha, alphaE, alphaE2;
michael@0:     float frame_loudness;
michael@0:     float bandwidth_mask;
michael@0:     int bandwidth=0;
michael@0:     float maxE = 0;
michael@0:     float noise_floor;
michael@0:     int remaining;
michael@0:     AnalysisInfo *info;
michael@0:     SAVE_STACK;
michael@0: 
michael@0:     tonal->last_transition++;
michael@0:     alpha = 1.f/IMIN(20, 1+tonal->count);
michael@0:     alphaE = 1.f/IMIN(50, 1+tonal->count);
michael@0:     alphaE2 = 1.f/IMIN(1000, 1+tonal->count);
michael@0: 
michael@0:     if (tonal->count<4)
michael@0:        tonal->music_prob = .5;
michael@0:     kfft = celt_mode->mdct.kfft[0];
michael@0:     if (tonal->count==0)
michael@0:        tonal->mem_fill = 240;
michael@0:     downmix(x, &tonal->inmem[tonal->mem_fill], IMIN(len, ANALYSIS_BUF_SIZE-tonal->mem_fill), offset, c1, c2, C);
michael@0:     if (tonal->mem_fill+len < ANALYSIS_BUF_SIZE)
michael@0:     {
michael@0:        tonal->mem_fill += len;
michael@0:        /* Don't have enough to update the analysis */
michael@0:        RESTORE_STACK;
michael@0:        return;
michael@0:     }
michael@0:     info = &tonal->info[tonal->write_pos++];
michael@0:     if (tonal->write_pos>=DETECT_SIZE)
michael@0:        tonal->write_pos-=DETECT_SIZE;
michael@0: 
michael@0:     ALLOC(in, 480, kiss_fft_cpx);
michael@0:     ALLOC(out, 480, kiss_fft_cpx);
michael@0:     ALLOC(tonality, 240, float);
michael@0:     ALLOC(noisiness, 240, float);
michael@0:     for (i=0;i<N2;i++)
michael@0:     {
michael@0:        float w = analysis_window[i];
michael@0:        in[i].r = (kiss_fft_scalar)(w*tonal->inmem[i]);
michael@0:        in[i].i = (kiss_fft_scalar)(w*tonal->inmem[N2+i]);
michael@0:        in[N-i-1].r = (kiss_fft_scalar)(w*tonal->inmem[N-i-1]);
michael@0:        in[N-i-1].i = (kiss_fft_scalar)(w*tonal->inmem[N+N2-i-1]);
michael@0:     }
michael@0:     OPUS_MOVE(tonal->inmem, tonal->inmem+ANALYSIS_BUF_SIZE-240, 240);
michael@0:     remaining = len - (ANALYSIS_BUF_SIZE-tonal->mem_fill);
michael@0:     downmix(x, &tonal->inmem[240], remaining, offset+ANALYSIS_BUF_SIZE-tonal->mem_fill, c1, c2, C);
michael@0:     tonal->mem_fill = 240 + remaining;
michael@0:     opus_fft(kfft, in, out);
michael@0: 
michael@0:     for (i=1;i<N2;i++)
michael@0:     {
michael@0:        float X1r, X2r, X1i, X2i;
michael@0:        float angle, d_angle, d2_angle;
michael@0:        float angle2, d_angle2, d2_angle2;
michael@0:        float mod1, mod2, avg_mod;
michael@0:        X1r = (float)out[i].r+out[N-i].r;
michael@0:        X1i = (float)out[i].i-out[N-i].i;
michael@0:        X2r = (float)out[i].i+out[N-i].i;
michael@0:        X2i = (float)out[N-i].r-out[i].r;
michael@0: 
michael@0:        angle = (float)(.5f/M_PI)*fast_atan2f(X1i, X1r);
michael@0:        d_angle = angle - A[i];
michael@0:        d2_angle = d_angle - dA[i];
michael@0: 
michael@0:        angle2 = (float)(.5f/M_PI)*fast_atan2f(X2i, X2r);
michael@0:        d_angle2 = angle2 - angle;
michael@0:        d2_angle2 = d_angle2 - d_angle;
michael@0: 
michael@0:        mod1 = d2_angle - (float)floor(.5+d2_angle);
michael@0:        noisiness[i] = ABS16(mod1);
michael@0:        mod1 *= mod1;
michael@0:        mod1 *= mod1;
michael@0: 
michael@0:        mod2 = d2_angle2 - (float)floor(.5+d2_angle2);
michael@0:        noisiness[i] += ABS16(mod2);
michael@0:        mod2 *= mod2;
michael@0:        mod2 *= mod2;
michael@0: 
michael@0:        avg_mod = .25f*(d2A[i]+2.f*mod1+mod2);
michael@0:        tonality[i] = 1.f/(1.f+40.f*16.f*pi4*avg_mod)-.015f;
michael@0: 
michael@0:        A[i] = angle2;
michael@0:        dA[i] = d_angle2;
michael@0:        d2A[i] = mod2;
michael@0:     }
michael@0: 
michael@0:     frame_tonality = 0;
michael@0:     max_frame_tonality = 0;
michael@0:     /*tw_sum = 0;*/
michael@0:     info->activity = 0;
michael@0:     frame_noisiness = 0;
michael@0:     frame_stationarity = 0;
michael@0:     if (!tonal->count)
michael@0:     {
michael@0:        for (b=0;b<NB_TBANDS;b++)
michael@0:        {
michael@0:           tonal->lowE[b] = 1e10;
michael@0:           tonal->highE[b] = -1e10;
michael@0:        }
michael@0:     }
michael@0:     relativeE = 0;
michael@0:     frame_loudness = 0;
michael@0:     for (b=0;b<NB_TBANDS;b++)
michael@0:     {
michael@0:        float E=0, tE=0, nE=0;
michael@0:        float L1, L2;
michael@0:        float stationarity;
michael@0:        for (i=tbands[b];i<tbands[b+1];i++)
michael@0:        {
michael@0:           float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
michael@0:                      + out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
michael@0: #ifdef FIXED_POINT
michael@0:           /* FIXME: It's probably best to change the BFCC filter initial state instead */
michael@0:           binE *= 5.55e-17f;
michael@0: #endif
michael@0:           E += binE;
michael@0:           tE += binE*tonality[i];
michael@0:           nE += binE*2.f*(.5f-noisiness[i]);
michael@0:        }
michael@0:        tonal->E[tonal->E_count][b] = E;
michael@0:        frame_noisiness += nE/(1e-15f+E);
michael@0: 
michael@0:        frame_loudness += (float)sqrt(E+1e-10f);
michael@0:        logE[b] = (float)log(E+1e-10f);
michael@0:        tonal->lowE[b] = MIN32(logE[b], tonal->lowE[b]+.01f);
michael@0:        tonal->highE[b] = MAX32(logE[b], tonal->highE[b]-.1f);
michael@0:        if (tonal->highE[b] < tonal->lowE[b]+1.f)
michael@0:        {
michael@0:           tonal->highE[b]+=.5f;
michael@0:           tonal->lowE[b]-=.5f;
michael@0:        }
michael@0:        relativeE += (logE[b]-tonal->lowE[b])/(1e-15f+tonal->highE[b]-tonal->lowE[b]);
michael@0: 
michael@0:        L1=L2=0;
michael@0:        for (i=0;i<NB_FRAMES;i++)
michael@0:        {
michael@0:           L1 += (float)sqrt(tonal->E[i][b]);
michael@0:           L2 += tonal->E[i][b];
michael@0:        }
michael@0: 
michael@0:        stationarity = MIN16(0.99f,L1/(float)sqrt(1e-15+NB_FRAMES*L2));
michael@0:        stationarity *= stationarity;
michael@0:        stationarity *= stationarity;
michael@0:        frame_stationarity += stationarity;
michael@0:        /*band_tonality[b] = tE/(1e-15+E)*/;
michael@0:        band_tonality[b] = MAX16(tE/(1e-15f+E), stationarity*tonal->prev_band_tonality[b]);
michael@0: #if 0
michael@0:        if (b>=NB_TONAL_SKIP_BANDS)
michael@0:        {
michael@0:           frame_tonality += tweight[b]*band_tonality[b];
michael@0:           tw_sum += tweight[b];
michael@0:        }
michael@0: #else
michael@0:        frame_tonality += band_tonality[b];
michael@0:        if (b>=NB_TBANDS-NB_TONAL_SKIP_BANDS)
michael@0:           frame_tonality -= band_tonality[b-NB_TBANDS+NB_TONAL_SKIP_BANDS];
michael@0: #endif
michael@0:        max_frame_tonality = MAX16(max_frame_tonality, (1.f+.03f*(b-NB_TBANDS))*frame_tonality);
michael@0:        slope += band_tonality[b]*(b-8);
michael@0:        /*printf("%f %f ", band_tonality[b], stationarity);*/
michael@0:        tonal->prev_band_tonality[b] = band_tonality[b];
michael@0:     }
michael@0: 
michael@0:     bandwidth_mask = 0;
michael@0:     bandwidth = 0;
michael@0:     maxE = 0;
michael@0:     noise_floor = 5.7e-4f/(1<<(IMAX(0,lsb_depth-8)));
michael@0: #ifdef FIXED_POINT
michael@0:     noise_floor *= 1<<(15+SIG_SHIFT);
michael@0: #endif
michael@0:     noise_floor *= noise_floor;
michael@0:     for (b=0;b<NB_TOT_BANDS;b++)
michael@0:     {
michael@0:        float E=0;
michael@0:        int band_start, band_end;
michael@0:        /* Keep a margin of 300 Hz for aliasing */
michael@0:        band_start = extra_bands[b];
michael@0:        band_end = extra_bands[b+1];
michael@0:        for (i=band_start;i<band_end;i++)
michael@0:        {
michael@0:           float binE = out[i].r*(float)out[i].r + out[N-i].r*(float)out[N-i].r
michael@0:                      + out[i].i*(float)out[i].i + out[N-i].i*(float)out[N-i].i;
michael@0:           E += binE;
michael@0:        }
michael@0:        maxE = MAX32(maxE, E);
michael@0:        tonal->meanE[b] = MAX32((1-alphaE2)*tonal->meanE[b], E);
michael@0:        E = MAX32(E, tonal->meanE[b]);
michael@0:        /* Use a simple follower with 13 dB/Bark slope for spreading function */
michael@0:        bandwidth_mask = MAX32(.05f*bandwidth_mask, E);
michael@0:        /* Consider the band "active" only if all these conditions are met:
michael@0:           1) less than 10 dB below the simple follower
michael@0:           2) less than 90 dB below the peak band (maximal masking possible considering
michael@0:              both the ATH and the loudness-dependent slope of the spreading function)
michael@0:           3) above the PCM quantization noise floor
michael@0:        */
michael@0:        if (E>.1*bandwidth_mask && E*1e9f > maxE && E > noise_floor*(band_end-band_start))
michael@0:           bandwidth = b;
michael@0:     }
michael@0:     if (tonal->count<=2)
michael@0:        bandwidth = 20;
michael@0:     frame_loudness = 20*(float)log10(frame_loudness);
michael@0:     tonal->Etracker = MAX32(tonal->Etracker-.03f, frame_loudness);
michael@0:     tonal->lowECount *= (1-alphaE);
michael@0:     if (frame_loudness < tonal->Etracker-30)
michael@0:        tonal->lowECount += alphaE;
michael@0: 
michael@0:     for (i=0;i<8;i++)
michael@0:     {
michael@0:        float sum=0;
michael@0:        for (b=0;b<16;b++)
michael@0:           sum += dct_table[i*16+b]*logE[b];
michael@0:        BFCC[i] = sum;
michael@0:     }
michael@0: 
michael@0:     frame_stationarity /= NB_TBANDS;
michael@0:     relativeE /= NB_TBANDS;
michael@0:     if (tonal->count<10)
michael@0:        relativeE = .5;
michael@0:     frame_noisiness /= NB_TBANDS;
michael@0: #if 1
michael@0:     info->activity = frame_noisiness + (1-frame_noisiness)*relativeE;
michael@0: #else
michael@0:     info->activity = .5*(1+frame_noisiness-frame_stationarity);
michael@0: #endif
michael@0:     frame_tonality = (max_frame_tonality/(NB_TBANDS-NB_TONAL_SKIP_BANDS));
michael@0:     frame_tonality = MAX16(frame_tonality, tonal->prev_tonality*.8f);
michael@0:     tonal->prev_tonality = frame_tonality;
michael@0: 
michael@0:     slope /= 8*8;
michael@0:     info->tonality_slope = slope;
michael@0: 
michael@0:     tonal->E_count = (tonal->E_count+1)%NB_FRAMES;
michael@0:     tonal->count++;
michael@0:     info->tonality = frame_tonality;
michael@0: 
michael@0:     for (i=0;i<4;i++)
michael@0:        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];
michael@0: 
michael@0:     for (i=0;i<4;i++)
michael@0:        tonal->cmean[i] = (1-alpha)*tonal->cmean[i] + alpha*BFCC[i];
michael@0: 
michael@0:     for (i=0;i<4;i++)
michael@0:         features[4+i] = 0.63246f*(BFCC[i]-tonal->mem[i+24]) + 0.31623f*(tonal->mem[i]-tonal->mem[i+16]);
michael@0:     for (i=0;i<3;i++)
michael@0:         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];
michael@0: 
michael@0:     if (tonal->count > 5)
michael@0:     {
michael@0:        for (i=0;i<9;i++)
michael@0:           tonal->std[i] = (1-alpha)*tonal->std[i] + alpha*features[i]*features[i];
michael@0:     }
michael@0: 
michael@0:     for (i=0;i<8;i++)
michael@0:     {
michael@0:        tonal->mem[i+24] = tonal->mem[i+16];
michael@0:        tonal->mem[i+16] = tonal->mem[i+8];
michael@0:        tonal->mem[i+8] = tonal->mem[i];
michael@0:        tonal->mem[i] = BFCC[i];
michael@0:     }
michael@0:     for (i=0;i<9;i++)
michael@0:        features[11+i] = (float)sqrt(tonal->std[i]);
michael@0:     features[20] = info->tonality;
michael@0:     features[21] = info->activity;
michael@0:     features[22] = frame_stationarity;
michael@0:     features[23] = info->tonality_slope;
michael@0:     features[24] = tonal->lowECount;
michael@0: 
michael@0: #ifndef DISABLE_FLOAT_API
michael@0:     mlp_process(&net, features, frame_probs);
michael@0:     frame_probs[0] = .5f*(frame_probs[0]+1);
michael@0:     /* Curve fitting between the MLP probability and the actual probability */
michael@0:     frame_probs[0] = .01f + 1.21f*frame_probs[0]*frame_probs[0] - .23f*(float)pow(frame_probs[0], 10);
michael@0:     /* Probability of active audio (as opposed to silence) */
michael@0:     frame_probs[1] = .5f*frame_probs[1]+.5f;
michael@0:     /* Consider that silence has a 50-50 probability. */
michael@0:     frame_probs[0] = frame_probs[1]*frame_probs[0] + (1-frame_probs[1])*.5f;
michael@0: 
michael@0:     /*printf("%f %f ", frame_probs[0], frame_probs[1]);*/
michael@0:     {
michael@0:        /* Probability of state transition */
michael@0:        float tau;
michael@0:        /* Represents independence of the MLP probabilities, where
michael@0:           beta=1 means fully independent. */
michael@0:        float beta;
michael@0:        /* Denormalized probability of speech (p0) and music (p1) after update */
michael@0:        float p0, p1;
michael@0:        /* Probabilities for "all speech" and "all music" */
michael@0:        float s0, m0;
michael@0:        /* Probability sum for renormalisation */
michael@0:        float psum;
michael@0:        /* Instantaneous probability of speech and music, with beta pre-applied. */
michael@0:        float speech0;
michael@0:        float music0;
michael@0: 
michael@0:        /* One transition every 3 minutes of active audio */
michael@0:        tau = .00005f*frame_probs[1];
michael@0:        beta = .05f;
michael@0:        if (1) {
michael@0:           /* Adapt beta based on how "unexpected" the new prob is */
michael@0:           float p, q;
michael@0:           p = MAX16(.05f,MIN16(.95f,frame_probs[0]));
michael@0:           q = MAX16(.05f,MIN16(.95f,tonal->music_prob));
michael@0:           beta = .01f+.05f*ABS16(p-q)/(p*(1-q)+q*(1-p));
michael@0:        }
michael@0:        /* p0 and p1 are the probabilities of speech and music at this frame
michael@0:           using only information from previous frame and applying the
michael@0:           state transition model */
michael@0:        p0 = (1-tonal->music_prob)*(1-tau) +    tonal->music_prob *tau;
michael@0:        p1 =    tonal->music_prob *(1-tau) + (1-tonal->music_prob)*tau;
michael@0:        /* We apply the current probability with exponent beta to work around
michael@0:           the fact that the probability estimates aren't independent. */
michael@0:        p0 *= (float)pow(1-frame_probs[0], beta);
michael@0:        p1 *= (float)pow(frame_probs[0], beta);
michael@0:        /* Normalise the probabilities to get the Marokv probability of music. */
michael@0:        tonal->music_prob = p1/(p0+p1);
michael@0:        info->music_prob = tonal->music_prob;
michael@0: 
michael@0:        /* This chunk of code deals with delayed decision. */
michael@0:        psum=1e-20f;
michael@0:        /* Instantaneous probability of speech and music, with beta pre-applied. */
michael@0:        speech0 = (float)pow(1-frame_probs[0], beta);
michael@0:        music0  = (float)pow(frame_probs[0], beta);
michael@0:        if (tonal->count==1)
michael@0:        {
michael@0:           tonal->pspeech[0]=.5;
michael@0:           tonal->pmusic [0]=.5;
michael@0:        }
michael@0:        /* Updated probability of having only speech (s0) or only music (m0),
michael@0:           before considering the new observation. */
michael@0:        s0 = tonal->pspeech[0] + tonal->pspeech[1];
michael@0:        m0 = tonal->pmusic [0] + tonal->pmusic [1];
michael@0:        /* Updates s0 and m0 with instantaneous probability. */
michael@0:        tonal->pspeech[0] = s0*(1-tau)*speech0;
michael@0:        tonal->pmusic [0] = m0*(1-tau)*music0;
michael@0:        /* Propagate the transition probabilities */
michael@0:        for (i=1;i<DETECT_SIZE-1;i++)
michael@0:        {
michael@0:           tonal->pspeech[i] = tonal->pspeech[i+1]*speech0;
michael@0:           tonal->pmusic [i] = tonal->pmusic [i+1]*music0;
michael@0:        }
michael@0:        /* Probability that the latest frame is speech, when all the previous ones were music. */
michael@0:        tonal->pspeech[DETECT_SIZE-1] = m0*tau*speech0;
michael@0:        /* Probability that the latest frame is music, when all the previous ones were speech. */
michael@0:        tonal->pmusic [DETECT_SIZE-1] = s0*tau*music0;
michael@0: 
michael@0:        /* Renormalise probabilities to 1 */
michael@0:        for (i=0;i<DETECT_SIZE;i++)
michael@0:           psum += tonal->pspeech[i] + tonal->pmusic[i];
michael@0:        psum = 1.f/psum;
michael@0:        for (i=0;i<DETECT_SIZE;i++)
michael@0:        {
michael@0:           tonal->pspeech[i] *= psum;
michael@0:           tonal->pmusic [i] *= psum;
michael@0:        }
michael@0:        psum = tonal->pmusic[0];
michael@0:        for (i=1;i<DETECT_SIZE;i++)
michael@0:           psum += tonal->pspeech[i];
michael@0: 
michael@0:        /* Estimate our confidence in the speech/music decisions */
michael@0:        if (frame_probs[1]>.75)
michael@0:        {
michael@0:           if (tonal->music_prob>.9)
michael@0:           {
michael@0:              float adapt;
michael@0:              adapt = 1.f/(++tonal->music_confidence_count);
michael@0:              tonal->music_confidence_count = IMIN(tonal->music_confidence_count, 500);
michael@0:              tonal->music_confidence += adapt*MAX16(-.2f,frame_probs[0]-tonal->music_confidence);
michael@0:           }
michael@0:           if (tonal->music_prob<.1)
michael@0:           {
michael@0:              float adapt;
michael@0:              adapt = 1.f/(++tonal->speech_confidence_count);
michael@0:              tonal->speech_confidence_count = IMIN(tonal->speech_confidence_count, 500);
michael@0:              tonal->speech_confidence += adapt*MIN16(.2f,frame_probs[0]-tonal->speech_confidence);
michael@0:           }
michael@0:        } else {
michael@0:           if (tonal->music_confidence_count==0)
michael@0:              tonal->music_confidence = .9f;
michael@0:           if (tonal->speech_confidence_count==0)
michael@0:              tonal->speech_confidence = .1f;
michael@0:        }
michael@0:     }
michael@0:     if (tonal->last_music != (tonal->music_prob>.5f))
michael@0:        tonal->last_transition=0;
michael@0:     tonal->last_music = tonal->music_prob>.5f;
michael@0: #else
michael@0:     info->music_prob = 0;
michael@0: #endif
michael@0:     /*for (i=0;i<25;i++)
michael@0:        printf("%f ", features[i]);
michael@0:     printf("\n");*/
michael@0: 
michael@0:     info->bandwidth = bandwidth;
michael@0:     /*printf("%d %d\n", info->bandwidth, info->opus_bandwidth);*/
michael@0:     info->noisiness = frame_noisiness;
michael@0:     info->valid = 1;
michael@0:     if (info_out!=NULL)
michael@0:        OPUS_COPY(info_out, info, 1);
michael@0:     RESTORE_STACK;
michael@0: }
michael@0: 
michael@0: void run_analysis(TonalityAnalysisState *analysis, const CELTMode *celt_mode, const void *analysis_pcm,
michael@0:                  int analysis_frame_size, int frame_size, int c1, int c2, int C, opus_int32 Fs,
michael@0:                  int lsb_depth, downmix_func downmix, AnalysisInfo *analysis_info)
michael@0: {
michael@0:    int offset;
michael@0:    int pcm_len;
michael@0: 
michael@0:    if (analysis_pcm != NULL)
michael@0:    {
michael@0:       /* Avoid overflow/wrap-around of the analysis buffer */
michael@0:       analysis_frame_size = IMIN((DETECT_SIZE-5)*Fs/100, analysis_frame_size);
michael@0: 
michael@0:       pcm_len = analysis_frame_size - analysis->analysis_offset;
michael@0:       offset = analysis->analysis_offset;
michael@0:       do {
michael@0:          tonality_analysis(analysis, NULL, celt_mode, analysis_pcm, IMIN(480, pcm_len), offset, c1, c2, C, lsb_depth, downmix);
michael@0:          offset += 480;
michael@0:          pcm_len -= 480;
michael@0:       } while (pcm_len>0);
michael@0:       analysis->analysis_offset = analysis_frame_size;
michael@0: 
michael@0:       analysis->analysis_offset -= frame_size;
michael@0:    }
michael@0: 
michael@0:    analysis_info->valid = 0;
michael@0:    tonality_get_info(analysis, analysis_info, frame_size);
michael@0: }