src/analysis.c

/* Copyright (c) 2011 Xiph.Org Foundation
   Written by Jean-Marc Valin */
/*
   Redistribution and use in source and binary forms, with or without
   modification, are permitted provided that the following conditions
   are met:

   - Redistributions of source code must retain the above copyright
   notice, this list of conditions and the following disclaimer.

   - Redistributions in binary form must reproduce the above copyright
   notice, this list of conditions and the following disclaimer in the
   documentation and/or other materials provided with the distribution.

   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
   A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
   CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#define ANALYSIS_C

#include <stdio.h>

#include "mathops.h"
#include "kiss_fft.h"
#include "celt.h"
#include "modes.h"
#include "arch.h"
#include "quant_bands.h"
#include "analysis.h"
#include "mlp.h"
#include "stack_alloc.h"
#include "float_cast.h"

#ifndef M_PI
#define M_PI 3.141592653
#endif

#ifndef DISABLE_FLOAT_API

#define TRANSITION_PENALTY 10

static const float dct_table[128] = {
        0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f,
        0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f, 0.250000f,
        0.351851f, 0.338330f, 0.311806f, 0.273300f, 0.224292f, 0.166664f, 0.102631f, 0.034654f,
       -0.034654f,-0.102631f,-0.166664f,-0.224292f,-0.273300f,-0.311806f,-0.338330f,-0.351851f,
        0.346760f, 0.293969f, 0.196424f, 0.068975f,-0.068975f,-0.196424f,-0.293969f,-0.346760f,
       -0.346760f,-0.293969f,-0.196424f,-0.068975f, 0.068975f, 0.196424f, 0.293969f, 0.346760f,
        0.338330f, 0.224292f, 0.034654f,-0.166664f,-0.311806f,-0.351851f,-0.273300f,-0.102631f,
        0.102631f, 0.273300f, 0.351851f, 0.311806f, 0.166664f,-0.034654f,-0.224292f,-0.338330f,
        0.326641f, 0.135299f,-0.135299f,-0.326641f,-0.326641f,-0.135299f, 0.135299f, 0.326641f,
        0.326641f, 0.135299f,-0.135299f,-0.326641f,-0.326641f,-0.135299f, 0.135299f, 0.326641f,
        0.311806f, 0.034654f,-0.273300f,-0.338330f,-0.102631f, 0.224292f, 0.351851f, 0.166664f,
       -0.166664f,-0.351851f,-0.224292f, 0.102631f, 0.338330f, 0.273300f,-0.034654f,-0.311806f,
        0.293969f,-0.068975f,-0.346760f,-0.196424f, 0.196424f, 0.346760f, 0.068975f,-0.293969f,
       -0.293969f, 0.068975f, 0.346760f, 0.196424f,-0.196424f,-0.346760f,-0.068975f, 0.293969f,
        0.273300f,-0.166664f,-0.338330f, 0.034654f, 0.351851f, 0.102631f,-0.311806f,-0.224292f,
        0.224292f, 0.311806f,-0.102631f,-0.351851f,-0.034654f, 0.338330f, 0.166664f,-0.273300f,
};

static const float analysis_window[240] = {
      0.000043f, 0.000171f, 0.000385f, 0.000685f, 0.001071f, 0.001541f, 0.002098f, 0.002739f,
      0.003466f, 0.004278f, 0.005174f, 0.006156f, 0.007222f, 0.008373f, 0.009607f, 0.010926f,
      0.012329f, 0.013815f, 0.015385f, 0.017037f, 0.018772f, 0.020590f, 0.022490f, 0.024472f,
      0.026535f, 0.028679f, 0.030904f, 0.033210f, 0.035595f, 0.038060f, 0.040604f, 0.043227f,
      0.045928f, 0.048707f, 0.051564f, 0.054497f, 0.057506f, 0.060591f, 0.063752f, 0.066987f,
      0.070297f, 0.073680f, 0.077136f, 0.080665f, 0.084265f, 0.087937f, 0.091679f, 0.095492f,
      0.099373f, 0.103323f, 0.107342f, 0.111427f, 0.115579f, 0.119797f, 0.124080f, 0.128428f,
      0.132839f, 0.137313f, 0.141849f, 0.146447f, 0.151105f, 0.155823f, 0.160600f, 0.165435f,
      0.170327f, 0.175276f, 0.180280f, 0.185340f, 0.190453f, 0.195619f, 0.200838f, 0.206107f,
      0.211427f, 0.216797f, 0.222215f, 0.227680f, 0.233193f, 0.238751f, 0.244353f, 0.250000f,
      0.255689f, 0.261421f, 0.267193f, 0.273005f, 0.278856f, 0.284744f, 0.290670f, 0.296632f,
      0.302628f, 0.308658f, 0.314721f, 0.320816f, 0.326941f, 0.333097f, 0.339280f, 0.345492f,
      0.351729f, 0.357992f, 0.364280f, 0.370590f, 0.376923f, 0.383277f, 0.389651f, 0.396044f,
      0.402455f, 0.408882f, 0.415325f, 0.421783f, 0.428254f, 0.434737f, 0.441231f, 0.447736f,
      0.454249f, 0.460770f, 0.467298f, 0.473832f, 0.480370f, 0.486912f, 0.493455f, 0.500000f,
      0.506545f, 0.513088f, 0.519630f, 0.526168f, 0.532702f, 0.539230f, 0.545751f, 0.552264f,
      0.558769f, 0.565263f, 0.571746f, 0.578217f, 0.584675f, 0.591118f, 0.597545f, 0.603956f,
      0.610349f, 0.616723f, 0.623077f, 0.629410f, 0.635720f, 0.642008f, 0.648271f, 0.654508f,
      0.660720f, 0.666903f, 0.673059f, 0.679184f, 0.685279f, 0.691342f, 0.697372f, 0.703368f,
      0.709330f, 0.715256f, 0.721144f, 0.726995f, 0.732807f, 0.738579f, 0.744311f, 0.750000f,
      0.755647f, 0.761249f, 0.766807f, 0.772320f, 0.777785f, 0.783203f, 0.788573f, 0.793893f,
      0.799162f, 0.804381f, 0.809547f, 0.814660f, 0.819720f, 0.824724f, 0.829673f, 0.834565f,
      0.839400f, 0.844177f, 0.848895f, 0.853553f, 0.858151f, 0.862687f, 0.867161f, 0.871572f,
      0.875920f, 0.880203f, 0.884421f, 0.888573f, 0.892658f, 0.896677f, 0.900627f, 0.904508f,
      0.908321f, 0.912063f, 0.915735f, 0.919335f, 0.922864f, 0.926320f, 0.929703f, 0.933013f,
      0.936248f, 0.939409f, 0.942494f, 0.945503f, 0.948436f, 0.951293f, 0.954072f, 0.956773f,
      0.959396f, 0.961940f, 0.964405f, 0.966790f, 0.969096f, 0.971321f, 0.973465f, 0.975528f,
      0.977510f, 0.979410f, 0.981228f, 0.982963f, 0.984615f, 0.986185f, 0.987671f, 0.989074f,
      0.990393f, 0.991627f, 0.992778f, 0.993844f, 0.994826f, 0.995722f, 0.996534f, 0.997261f,
      0.997902f, 0.998459f, 0.998929f, 0.999315f, 0.999615f, 0.999829f, 0.999957f, 1.000000f,
};

static const int tbands[NB_TBANDS+1] = {
      4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 80, 96, 112, 136, 160, 192, 240
};

#define NB_TONAL_SKIP_BANDS 9

static opus_val32 silk_resampler_down2_hp(
    opus_val32                  *S,                 /* I/O  State vector [ 2 ]                                          */
    opus_val32                  *out,               /* O    Output signal [ floor(len/2) ]                              */
    const opus_val32            *in,                /* I    Input signal [ len ]                                        */
    int                         inLen               /* I    Number of input samples                                     */
)
{
    int k, len2 = inLen/2;
    opus_val32 in32, out32, out32_hp, Y, X;
    opus_val64 hp_ener = 0;
    /* Internal variables and state are in Q10 format */
    for( k = 0; k < len2; k++ ) {
        /* Convert to Q10 */
        in32 = in[ 2 * k ];

        /* All-pass section for even input sample */
        Y      = SUB32( in32, S[ 0 ] );
        X      = MULT16_32_Q15(QCONST16(0.6074371f, 15), Y);
        out32  = ADD32( S[ 0 ], X );
        S[ 0 ] = ADD32( in32, X );
        out32_hp = out32;
        /* Convert to Q10 */
        in32 = in[ 2 * k + 1 ];

        /* All-pass section for odd input sample, and add to output of previous section */
        Y      = SUB32( in32, S[ 1 ] );
        X      = MULT16_32_Q15(QCONST16(0.15063f, 15), Y);
        out32  = ADD32( out32, S[ 1 ] );
        out32  = ADD32( out32, X );
        S[ 1 ] = ADD32( in32, X );

        Y      = SUB32( -in32, S[ 2 ] );
        X      = MULT16_32_Q15(QCONST16(0.15063f, 15), Y);
        out32_hp  = ADD32( out32_hp, S[ 2 ] );
        out32_hp  = ADD32( out32_hp, X );
        S[ 2 ] = ADD32( -in32, X );

        hp_ener += out32_hp*(opus_val64)out32_hp;
        /* Add, convert back to int16 and store to output */
        out[ k ] = HALF32(out32);
    }
#ifdef FIXED_POINT
    /* len2 can be up to 480, so we shift by 8 more to make it fit. */
    hp_ener = hp_ener >> (2*SIG_SHIFT + 8);
#endif
    return (opus_val32)hp_ener;
}

static opus_val32 downmix_and_resample(downmix_func downmix, const void *_x, opus_val32 *y, opus_val32 S[3], int subframe, int offset, int c1, int c2, int C, int Fs)
{
   VARDECL(opus_val32, tmp);
   opus_val32 scale;
   int j;
   opus_val32 ret = 0;
   SAVE_STACK;

   if (subframe==0) return 0;
   if (Fs == 48000)
   {
      subframe *= 2;
      offset *= 2;
   } else if (Fs == 16000) {
      subframe = subframe*2/3;
      offset = offset*2/3;
   }
   ALLOC(tmp, subframe, opus_val32);

   downmix(_x, tmp, subframe, offset, c1, c2, C);
#ifdef FIXED_POINT
   scale = (1<<SIG_SHIFT);
#else
   scale = 1.f/32768;
#endif
   if (c2==-2)
      scale /= C;
   else if (c2>-1)
      scale /= 2;
   for (j=0;j<subframe;j++)
      tmp[j] *= scale;
   if (Fs == 48000)
   {
      ret = silk_resampler_down2_hp(S, y, tmp, subframe);
   } else if (Fs == 24000) {
      OPUS_COPY(y, tmp, subframe);
   } else if (Fs == 16000) {
      VARDECL(opus_val32, tmp3x);
      ALLOC(tmp3x, 3*subframe, opus_val32);
      /* Don't do this at home! This resampler is horrible and it's only (barely)
         usable for the purpose of the analysis because we don't care about all
         the aliasing between 8 kHz and 12 kHz. */
      for (j=0;j<subframe;j++)
      {
         tmp3x[3*j] = tmp[j];
         tmp3x[3*j+1] = tmp[j];
         tmp3x[3*j+2] = tmp[j];
      }
      silk_resampler_down2_hp(S, y, tmp3x, 3*subframe);
   }
   RESTORE_STACK;
   return ret;
}

void tonality_analysis_init(TonalityAnalysisState *tonal, opus_int32 Fs)
{
  /* Initialize reusable fields. */
  tonal->arch = opus_select_arch();
  tonal->Fs = Fs;
  /* Clear remaining fields. */
  tonality_analysis_reset(tonal);
}

void tonality_analysis_reset(TonalityAnalysisState *tonal)
{
  /* Clear non-reusable fields. */
  char *start = (char*)&tonal->TONALITY_ANALYSIS_RESET_START;
  OPUS_CLEAR(start, sizeof(TonalityAnalysisState) - (start - (char*)tonal));
}

void tonality_get_info(TonalityAnalysisState *tonal, AnalysisInfo *info_out, int len)
{
   int pos;
   int curr_lookahead;
   float tonality_max;
   float tonality_avg;
   int tonality_count;
   int i;
   int pos0;
   float prob_avg;
   float prob_count;
   float prob_min, prob_max;
   float vad_prob;
   int mpos, vpos;

   pos = tonal->read_pos;
   curr_lookahead = tonal->write_pos-tonal->read_pos;
   if (curr_lookahead<0)
      curr_lookahead += DETECT_SIZE;

   /* On long frames, look at the second analysis window rather than the first. */
   if (len > tonal->Fs/50 && pos != tonal->write_pos)
   {
      pos++;
      if (pos==DETECT_SIZE)
         pos=0;
   }
   if (pos == tonal->write_pos)
      pos--;
   if (pos<0)
      pos = DETECT_SIZE-1;
   pos0 = pos;
   OPUS_COPY(info_out, &tonal->info[pos], 1);
   tonality_max = tonality_avg = info_out->tonality;
   tonality_count = 1;
   /* If possible, look ahead for a tone to compensate for the delay in the tone detector. */
   for (i=0;i<3;i++)
   {
      pos++;
      if (pos==DETECT_SIZE)
         pos = 0;
      if (pos == tonal->write_pos)
         break;
      tonality_max = MAX32(tonality_max, tonal->info[pos].tonality);
      tonality_avg += tonal->info[pos].tonality;
      tonality_count++;
   }
   info_out->tonality = MAX32(tonality_avg/tonality_count, tonality_max-.2f);

   mpos = vpos = pos0;
   /* If we have enough look-ahead, compensate for the ~5-frame delay in the music prob and
      ~1 frame delay in the VAD prob. */
   if (curr_lookahead > 15)
   {
      mpos += 5;
      if (mpos>=DETECT_SIZE)
         mpos -= DETECT_SIZE;
      vpos += 1;
      if (vpos>=DETECT_SIZE)
         vpos -= DETECT_SIZE;
   }

   /* The following calculations attempt to minimize a "badness function"
      for the transition. When switching from speech to music, the badness
      of switching at frame k is
      b_k = S*v_k + \sum_{i=0}^{k-1} v_i*(p_i - T)
      where
      v_i is the activity probability (VAD) at frame i,
      p_i is the music probability at frame i
      T is the probability threshold for switching
      S is the penalty for switching during active audio rather than silence
      the current frame has index i=0

      Rather than apply badness to directly decide when to switch, what we compute
      instead is the threshold for which the optimal switching point is now. When
      considering whether to switch now (frame 0) or at frame k, we have:
      S*v_0 = S*v_k + \sum_{i=0}^{k-1} v_i*(p_i - T)
      which gives us:
      T = ( \sum_{i=0}^{k-1} v_i*p_i + S*(v_k-v_0) ) / ( \sum_{i=0}^{k-1} v_i )
      We take the min threshold across all positive values of k (up to the maximum
      amount of lookahead we have) to give us the threshold for which the current
      frame is the optimal switch point.

      The last step is that we need to consider whether we want to switch at all.
      For that we use the average of the music probability over the entire window.
      If the threshold is higher than that average we're not going to
      switch, so we compute a min with the average as well. The result of all these
      min operations is music_prob_min, which gives the threshold for switching to music
      if we're currently encoding for speech.

      We do the exact opposite to compute music_prob_max which is used for switching
      from music to speech.
    */
   prob_min = 1.f;
   prob_max = 0.f;
   vad_prob = tonal->info[vpos].activity_probability;
   prob_count = MAX16(.1f, vad_prob);
   prob_avg = MAX16(.1f, vad_prob)*tonal->info[mpos].music_prob;
   while (1)
   {
      float pos_vad;
      mpos++;
      if (mpos==DETECT_SIZE)
         mpos = 0;
      if (mpos == tonal->write_pos)
         break;
      vpos++;
      if (vpos==DETECT_SIZE)
         vpos = 0;
      if (vpos == tonal->write_pos)
         break;
      pos_vad = tonal->info[vpos].activity_probability;
      prob_min = MIN16((prob_avg - TRANSITION_PENALTY*(vad_prob - pos_vad))/prob_count, prob_min);
      prob_max = MAX16((prob_avg + TRANSITION_PENALTY*(vad_prob - pos_vad))/prob_count, prob_max);
      prob_count += MAX16(.1f, pos_vad);
      prob_avg += MAX16(.1f, pos_vad)*tonal->info[mpos].music_prob;
   }
   info_out->music_prob = prob_avg/prob_count;
   prob_min = MIN16(prob_avg/prob_count, prob_min);
   prob_max = MAX16(prob_avg/prob_count, prob_max);
   prob_min = MAX16(prob_min, 0.f);
   prob_max = MIN16(prob_max, 1.f);

   /* If we don't have enough look-ahead, do our best to make a decent decision. */
   if (curr_lookahead < 10)
   {
      float pmin, pmax;
      pmin = prob_min;
      pmax = prob_max;
      pos = pos0;
      /* Look for min/max in the past. */
      for (i=0;i<IMIN(tonal->count-1, 15);i++)
      {
         pos--;
         if (pos < 0)
            pos = DETECT_SIZE-1;
         pmin = MIN16(pmin, tonal->info[pos].music_prob);
         pmax = MAX16(pmax, tonal->info[pos].music_prob);
      }
      /* Bias against switching on active audio. */
      pmin = MAX16(0.f, pmin - .1f*vad_prob);
      pmax = MIN16(1.f, pmax + .1f*vad_prob);
      prob_min += (1.f-.1f*curr_lookahead)*(pmin - prob_min);
      prob_max += (1.f-.1f*curr_lookahead)*(pmax - prob_max);
   }
   info_out->music_prob_min = prob_min;
   info_out->music_prob_max = prob_max;

   /* printf("%f %f %f %f %f\n", prob_min, prob_max, prob_avg/prob_count, vad_prob, info_out->music_prob); */
   tonal->read_subframe += len/(tonal->Fs/400);
   while (tonal->read_subframe>=8)
   {
      tonal->read_subframe -= 8;
      tonal->read_pos++;
   }
   if (tonal->read_pos>=DETECT_SIZE)
      tonal->read_pos-=DETECT_SIZE;
}

static const float std_feature_bias[9] = {
      5.684947f, 3.475288f, 1.770634f, 1.599784f, 3.773215f,
      2.163313f, 1.260756f, 1.116868f, 1.918795f
};

#define LEAKAGE_OFFSET 2.5f
#define LEAKAGE_SLOPE 2.f

#ifdef FIXED_POINT
/* For fixed-point, the input is +/-2^15 shifted up by SIG_SHIFT, so we need to
   compensate for that in the energy. */
#define SCALE_COMPENS (1.f/((opus_int32)1<<(15+SIG_SHIFT)))
#define SCALE_ENER(e) ((SCALE_COMPENS*SCALE_COMPENS)*(e))
#else
#define SCALE_ENER(e) (e)
#endif

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 is_masked[NB_TBANDS+1];
    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];
    float layer_out[MAX_NEURONS];
    SAVE_STACK;

    alpha = 1.f/IMIN(10, 1+tonal->count);
    alphaE = 1.f/IMIN(25, 1+tonal->count);
    /* Noise floor related decay for bandwidth detection: -2.2 dB/second */
    alphaE2 = 1.f/IMIN(100, 1+tonal->count);
    if (tonal->count <= 1) alphaE2 = 1;

    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;
    }

    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;
       float Em;
       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);
       Em = MAX32(E, tonal->meanE[b]);
       /* Consider the band "active" only if all these conditions are met:
          1) less than 90 dB below the peak band (maximal masking possible considering
             both the ATH and the loudness-dependent slope of the spreading function)
          2) above the PCM quantization noise floor
          We use b+1 because the first CELT band isn't included in tbands[]
       */
       if (E*1e9f > maxE && (Em > 3*noise_floor*(band_end-band_start) || E > noise_floor*(band_end-band_start)))
          bandwidth = b+1;
       /* Check if the band is masked (see below). */
       is_masked[b] = E < (tonal->prev_bandwidth >= b+1  ? .01f : .05f)*bandwidth_mask;
       /* Use a simple follower with 13 dB/Bark slope for spreading function. */
       bandwidth_mask = MAX32(.05f*bandwidth_mask, E);
    }
    /* Special case for the last two bands, for which we don't have spectrum but only
       the energy above 12 kHz. The difficulty here is that the high-pass we use
       leaks some LF energy, so we need to increase the threshold without accidentally cutting
       off the band. */
    if (tonal->Fs == 48000) {
       float noise_ratio;
       float Em;
       float E = hp_ener*(1.f/(60*60));
       noise_ratio = tonal->prev_bandwidth==20 ? 10.f : 30.f;

#ifdef FIXED_POINT
       /* silk_resampler_down2_hp() shifted right by an extra 8 bits. */
       E *= 256.f*(1.f/Q15ONE)*(1.f/Q15ONE);
#endif
       tonal->meanE[b] = MAX32((1-alphaE2)*tonal->meanE[b], E);
       Em = MAX32(E, tonal->meanE[b]);
       if (Em > 3*noise_ratio*noise_floor*160 || E > noise_ratio*noise_floor*160)
          bandwidth = 20;
       /* Check if the band is masked (see below). */
       is_masked[b] = E < (tonal->prev_bandwidth == 20  ? .01f : .05f)*bandwidth_mask;
    }
    /* In some cases, resampling aliasing can create a small amount of energy in the first band
       being cut. So if the last band is masked, we don't include it.  */
    if (bandwidth == 20 && is_masked[NB_TBANDS])
       bandwidth-=2;
    else if (bandwidth > 0 && bandwidth <= NB_TBANDS && is_masked[bandwidth-1])
       bandwidth--;
    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;

    compute_dense(&layer0, layer_out, features);
    compute_gru(&layer1, tonal->rnn_state, layer_out);
    compute_dense(&layer2, frame_probs, tonal->rnn_state);

    /* Probability of speech or music vs noise */
    info->activity_probability = frame_probs[1];
    /* It seems like the RNN tends to have a bias towards speech and this
       warping of the probabilities compensates for it. */
    info->music_prob = MAX16(1-10*(1-frame_probs[0]), MIN16(10*frame_probs[0], .12+.69*frame_probs[0]*(2-frame_probs[0])));

    /*printf("%f %f %f\n", frame_probs[0], frame_probs[1], info->music_prob);*/
#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;
}

void run_analysis(TonalityAnalysisState *analysis, const CELTMode *celt_mode, const void *analysis_pcm,
                 int analysis_frame_size, int frame_size, int c1, int c2, int C, opus_int32 Fs,
                 int lsb_depth, downmix_func downmix, AnalysisInfo *analysis_info)
{
   int offset;
   int pcm_len;

   analysis_frame_size -= analysis_frame_size&1;
   if (analysis_pcm != NULL)
   {
      /* Avoid overflow/wrap-around of the analysis buffer */
      analysis_frame_size = IMIN((DETECT_SIZE-5)*Fs/50, analysis_frame_size);

      pcm_len = analysis_frame_size - analysis->analysis_offset;
      offset = analysis->analysis_offset;
      while (pcm_len>0) {
         tonality_analysis(analysis, celt_mode, analysis_pcm, IMIN(Fs/50, pcm_len), offset, c1, c2, C, lsb_depth, downmix);
         offset += Fs/50;
         pcm_len -= Fs/50;
      }
      analysis->analysis_offset = analysis_frame_size;

      analysis->analysis_offset -= frame_size;
   }

   analysis_info->valid = 0;
   tonality_get_info(analysis, analysis_info, frame_size);
}

#endif /* DISABLE_FLOAT_API */