---

   /* FIXME: MC conversion required */
   /* Update weight to prevent circular convolution (MDF / AUMDF) */
   for (chan = 0; chan < C; chan++)
   {
      for (speak = 0; speak < K; speak++)
      {
         for (j=0;j<M;j++)
         {
            /* This is a variant of the Alternatively Updated MDF (AUMDF) */
            /* Remove the "if" to make this an MDF filter */
            if (j==0 || st->cancel_count%(M-1) == j-1)
            {
#ifdef FIXED_POINT
               for (i=0;i<N;i++)
                  st->wtmp2[i] = EXTRACT16(PSHR32(st->W[chan*N*K*M + j*N*K + speak*N + i],NORMALIZE_SCALEDOWN+16));
               spx_ifft(st->fft_table, st->wtmp2, st->wtmp);
               for (i=0;i<st->frame_size;i++)
               {
                  st->wtmp[i]=0;
               }
               for (i=st->frame_size;i<N;i++)
               {
                  st->wtmp[i]=SHL16(st->wtmp[i],NORMALIZE_SCALEUP);
               }
               spx_fft(st->fft_table, st->wtmp, st->wtmp2);
               /* The "-1" in the shift is a sort of kludge that trades less efficient update speed for decrease noise */
               for (i=0;i<N;i++)
                  st->W[chan*N*K*M + j*N*K + speak*N + i] -= SHL32(EXTEND32(st->wtmp2[i]),16+NORMALIZE_SCALEDOWN-NORMALIZE_SCALEUP-1);
#else
               spx_ifft(st->fft_table, &st->W[chan*N*K*M + j*N*K + speak*N], st->wtmp);
               for (i=st->frame_size;i<N;i++)
               {
                  st->wtmp[i]=0;
               }
               spx_fft(st->fft_table, st->wtmp, &st->W[chan*N*K*M + j*N*K + speak*N]);
#endif
            }
         }
      }
   }

   /* So we can use power_spectrum_accum */
   for (i=0;i<=st->frame_size;i++)
      st->Rf[i] = st->Yf[i] = st->Xf[i] = 0;

   Dbf = 0;
   See = 0;
#ifdef TWO_PATH
   /* Difference in response, this is used to estimate the variance of our residual power estimate */
   for (chan = 0; chan < C; chan++)
   {
      spectral_mul_accum(st->X, st->W+chan*N*K*M, st->Y+chan*N, N, M*K);
      spx_ifft(st->fft_table, st->Y+chan*N, st->y+chan*N);
      for (i=0;i<st->frame_size;i++)
         st->e[chan*N+i] = SUB16(st->e[chan*N+i+st->frame_size], st->y[chan*N+i+st->frame_size]);
      Dbf += 10+mdf_inner_prod(st->e+chan*N, st->e+chan*N, st->frame_size);
      for (i=0;i<st->frame_size;i++)
         st->e[chan*N+i] = SUB16(st->input[chan*st->frame_size+i], st->y[chan*N+i+st->frame_size]);
      See += mdf_inner_prod(st->e+chan*N, st->e+chan*N, st->frame_size);
   }
#endif

#ifndef TWO_PATH
   Sff = See;
#endif

#ifdef TWO_PATH
   /* Logic for updating the foreground filter */

   /* For two time windows, compute the mean of the energy difference, as well as the variance */
   st->Davg1 = ADD32(MULT16_32_Q15(QCONST16(.6f,15),st->Davg1), MULT16_32_Q15(QCONST16(.4f,15),SUB32(Sff,See)));
   st->Davg2 = ADD32(MULT16_32_Q15(QCONST16(.85f,15),st->Davg2), MULT16_32_Q15(QCONST16(.15f,15),SUB32(Sff,See)));
   st->Dvar1 = FLOAT_ADD(FLOAT_MULT(VAR1_SMOOTH, st->Dvar1), FLOAT_MUL32U(MULT16_32_Q15(QCONST16(.4f,15),Sff), MULT16_32_Q15(QCONST16(.4f,15),Dbf)));
   st->Dvar2 = FLOAT_ADD(FLOAT_MULT(VAR2_SMOOTH, st->Dvar2), FLOAT_MUL32U(MULT16_32_Q15(QCONST16(.15f,15),Sff), MULT16_32_Q15(QCONST16(.15f,15),Dbf)));

   /* Equivalent float code:
   st->Davg1 = .6*st->Davg1 + .4*(Sff-See);
   st->Davg2 = .85*st->Davg2 + .15*(Sff-See);
   st->Dvar1 = .36*st->Dvar1 + .16*Sff*Dbf;
   st->Dvar2 = .7225*st->Dvar2 + .0225*Sff*Dbf;
   */

   update_foreground = 0;
   /* Check if we have a statistically significant reduction in the residual echo */
   /* Note that this is *not* Gaussian, so we need to be careful about the longer tail */
   if (FLOAT_GT(FLOAT_MUL32U(SUB32(Sff,See),ABS32(SUB32(Sff,See))), FLOAT_MUL32U(Sff,Dbf)))
      update_foreground = 1;
   else if (FLOAT_GT(FLOAT_MUL32U(st->Davg1, ABS32(st->Davg1)), FLOAT_MULT(VAR1_UPDATE,(st->Dvar1))))
      update_foreground = 1;
   else if (FLOAT_GT(FLOAT_MUL32U(st->Davg2, ABS32(st->Davg2)), FLOAT_MULT(VAR2_UPDATE,(st->Dvar2))))
      update_foreground = 1;

   /* Do we update? */
   if (update_foreground)
   {
      st->Davg1 = st->Davg2 = 0;
      st->Dvar1 = st->Dvar2 = FLOAT_ZERO;
      /* Copy background filter to foreground filter */
      for (i=0;i<N*M*C*K;i++)
         st->foreground[i] = EXTRACT16(PSHR32(st->W[i],16));
      /* Apply a smooth transition so as to not introduce blocking artifacts */
      for (chan = 0; chan < C; chan++)
         for (i=0;i<st->frame_size;i++)
            st->e[chan*N+i+st->frame_size] = MULT16_16_Q15(st->window[i+st->frame_size],st->e[chan*N+i+st->frame_size]) + MULT16_16_Q15(st->window[i],st->y[chan*N+i+st->frame_size]);
   } else {
      int reset_background=0;
      /* Otherwise, check if the background filter is significantly worse */
      if (FLOAT_GT(FLOAT_MUL32U(NEG32(SUB32(Sff,See)),ABS32(SUB32(Sff,See))), FLOAT_MULT(VAR_BACKTRACK,FLOAT_MUL32U(Sff,Dbf))))
         reset_background = 1;
      if (FLOAT_GT(FLOAT_MUL32U(NEG32(st->Davg1), ABS32(st->Davg1)), FLOAT_MULT(VAR_BACKTRACK,st->Dvar1)))
         reset_background = 1;
      if (FLOAT_GT(FLOAT_MUL32U(NEG32(st->Davg2), ABS32(st->Davg2)), FLOAT_MULT(VAR_BACKTRACK,st->Dvar2)))
         reset_background = 1;
      if (reset_background)
      {
         /* Copy foreground filter to background filter */
         for (i=0;i<N*M*C*K;i++)
            st->W[i] = SHL32(EXTEND32(st->foreground[i]),16);
         /* We also need to copy the output so as to get correct adaptation */
         for (chan = 0; chan < C; chan++)
         {
            for (i=0;i<st->frame_size;i++)
               st->y[chan*N+i+st->frame_size] = st->e[chan*N+i+st->frame_size];
            for (i=0;i<st->frame_size;i++)
               st->e[chan*N+i] = SUB16(st->input[chan*st->frame_size+i], st->y[chan*N+i+st->frame_size]);
         }
         See = Sff;
         st->Davg1 = st->Davg2 = 0;
         st->Dvar1 = st->Dvar2 = FLOAT_ZERO;
      }
   }
#endif

   Sey = Syy = Sdd = 0;
   for (chan = 0; chan < C; chan++)
   {
      /* Compute error signal (for the output with de-emphasis) */
      for (i=0;i<st->frame_size;i++)
      {
         spx_word32_t tmp_out;
#ifdef TWO_PATH
         tmp_out = SUB32(EXTEND32(st->input[chan*st->frame_size+i]), EXTEND32(st->e[chan*N+i+st->frame_size]));
#else
         tmp_out = SUB32(EXTEND32(st->input[chan*st->frame_size+i]), EXTEND32(st->y[chan*N+i+st->frame_size]));
#endif
         tmp_out = ADD32(tmp_out, EXTEND32(MULT16_16_P15(st->preemph, st->memE[chan])));
      /* This is an arbitrary test for saturation in the microphone signal */
         if (in[i*C+chan] <= -32000 || in[i*C+chan] >= 32000)
         {
         if (st->saturated == 0)
            st->saturated = 1;
         }
         out[i*C+chan] = WORD2INT(tmp_out);
         st->memE[chan] = tmp_out;
      }

#ifdef DUMP_ECHO_CANCEL_DATA
      dump_audio(in, far_end, out, st->frame_size);
#endif

      /* Compute error signal (filter update version) */
      for (i=0;i<st->frame_size;i++)
      {
         st->e[chan*N+i+st->frame_size] = st->e[chan*N+i];
         st->e[chan*N+i] = 0;
      }

      /* Compute a bunch of correlations */
      /* FIXME: bad merge */
      Sey += mdf_inner_prod(st->e+chan*N+st->frame_size, st->y+chan*N+st->frame_size, st->frame_size);
      Syy += mdf_inner_prod(st->y+chan*N+st->frame_size, st->y+chan*N+st->frame_size, st->frame_size);
      Sdd += mdf_inner_prod(st->input+chan*st->frame_size, st->input+chan*st->frame_size, st->frame_size);

      /* Convert error to frequency domain */
      spx_fft(st->fft_table, st->e+chan*N, st->E+chan*N);
      for (i=0;i<st->frame_size;i++)
         st->y[i+chan*N] = 0;
      spx_fft(st->fft_table, st->y+chan*N, st->Y+chan*N);

      /* Compute power spectrum of echo (X), error (E) and filter response (Y) */
      power_spectrum_accum(st->E+chan*N, st->Rf, N);
      power_spectrum_accum(st->Y+chan*N, st->Yf, N);

   }

   /*printf ("%f %f %f %f\n", Sff, See, Syy, Sdd, st->update_cond);*/

   /* Do some sanity check */
   if (!(Syy>=0 && Sxx>=0 && See >= 0)
#ifndef FIXED_POINT
       || !(Sff < N*1e9 && Syy < N*1e9 && Sxx < N*1e9)
#endif
      )
   {
      /* Things have gone really bad */
      st->screwed_up += 50;
      for (i=0;i<st->frame_size*C;i++)
         out[i] = 0;
   } else if (SHR32(Sff, 2) > ADD32(Sdd, SHR32(MULT16_16(N, 10000),6)))
   {
      /* AEC seems to add lots of echo instead of removing it, let's see if it will improve */
      st->screwed_up++;
   } else {
      /* Everything's fine */
      st->screwed_up=0;
   }
   if (st->screwed_up>=50)
   {
      speex_warning("The echo canceller started acting funny and got slapped (reset). It swears it will behave now.");
      speex_echo_state_reset(st);
      return;
   }

   /* Add a small noise floor to make sure not to have problems when dividing */
   See = MAX32(See, SHR32(MULT16_16(N, 100),6));

   for (speak = 0; speak < K; speak++)
   {
      Sxx += mdf_inner_prod(st->x+speak*N+st->frame_size, st->x+speak*N+st->frame_size, st->frame_size);
      power_spectrum_accum(st->X+speak*N, st->Xf, N);
   }


   /* Smooth far end energy estimate over time */
   for (j=0;j<=st->frame_size;j++)
      st->power[j] = MULT16_32_Q15(ss_1,st->power[j]) + 1 + MULT16_32_Q15(ss,st->Xf[j]);

   /* Compute filtered spectra and (cross-)correlations */
   for (j=st->frame_size;j>=0;j--)
   {
      spx_float_t Eh, Yh;
      Eh = PSEUDOFLOAT(st->Rf[j] - st->Eh[j]);
      Yh = PSEUDOFLOAT(st->Yf[j] - st->Yh[j]);
      Pey = FLOAT_ADD(Pey,FLOAT_MULT(Eh,Yh));
      Pyy = FLOAT_ADD(Pyy,FLOAT_MULT(Yh,Yh));
#ifdef FIXED_POINT
      st->Eh[j] = MAC16_32_Q15(MULT16_32_Q15(SUB16(32767,st->spec_average),st->Eh[j]), st->spec_average, st->Rf[j]);
      st->Yh[j] = MAC16_32_Q15(MULT16_32_Q15(SUB16(32767,st->spec_average),st->Yh[j]), st->spec_average, st->Yf[j]);
#else
      st->Eh[j] = (1-st->spec_average)*st->Eh[j] + st->spec_average*st->Rf[j];
      st->Yh[j] = (1-st->spec_average)*st->Yh[j] + st->spec_average*st->Yf[j];
#endif
   }