diff --git a/libcelt/bands.c b/libcelt/bands.c
index 7db4bad7..aef14651 100644
--- a/libcelt/bands.c
+++ b/libcelt/bands.c
@@ -91,50 +91,71 @@ static void exp_rotation(float *X, int len, float theta, int dir, int stride, in
 /* Compute the amplitude (sqrt energy) in each of the bands */
 void compute_band_energies(const CELTMode *m, float *X, float *bank)
 {
-   int i, B;
+   int i, c, B, C;
    const int *eBands = m->eBands;
-   B = m->nbMdctBlocks*m->nbChannels;
-   for (i=0;i<m->nbEBands;i++)
+   B = m->nbMdctBlocks;
+   C = m->nbChannels;
+   for (c=0;c<C;c++)
    {
-      int j;
-      bank[i] = 1e-10;
-      for (j=B*eBands[i];j<B*eBands[i+1];j++)
-         bank[i] += X[j]*X[j];
-      bank[i] = sqrt(bank[i]);
+      for (i=0;i<m->nbEBands;i++)
+      {
+         int j;
+         float sum = 1e-10;
+         for (j=B*eBands[i];j<B*eBands[i+1];j++)
+            sum += X[j*C+c]*X[j*C+c];
+         bank[i*C+c] = sqrt(C*sum);
+         //printf ("%f ", bank[i*C+c]);
+      }
    }
+   //printf ("\n");
 }
 
 /* Normalise each band such that the energy is one. */
 void normalise_bands(const CELTMode *m, float *X, float *bank)
 {
-   int i, B;
+   int i, c, B, C;
    const int *eBands = m->eBands;
-   B = m->nbMdctBlocks*m->nbChannels;
-   for (i=0;i<m->nbEBands;i++)
+   B = m->nbMdctBlocks;
+   C = m->nbChannels;
+   for (c=0;c<C;c++)
    {
-      int j;
-      float x = 1.f/(1e-10+bank[i]);
-      for (j=B*eBands[i];j<B*eBands[i+1];j++)
-         X[j] *= x;
+      for (i=0;i<m->nbEBands;i++)
+      {
+         int j;
+         float g = 1.f/(1e-10+bank[i*C+c]);
+         for (j=B*eBands[i];j<B*eBands[i+1];j++)
+            X[j*C+c] *= g;
+      }
    }
-   for (i=B*eBands[m->nbEBands];i<B*eBands[m->nbEBands+1];i++)
+   for (i=B*C*eBands[m->nbEBands];i<B*C*eBands[m->nbEBands+1];i++)
       X[i] = 0;
 }
 
+void renormalise_bands(const CELTMode *m, float *X)
+{
+   float tmpE[m->nbEBands*m->nbChannels];
+   compute_band_energies(m, X, tmpE);
+   normalise_bands(m, X, tmpE);
+}
+
 /* De-normalise the energy to produce the synthesis from the unit-energy bands */
 void denormalise_bands(const CELTMode *m, float *X, float *bank)
 {
-   int i, B;
+   int i, c, B, C;
    const int *eBands = m->eBands;
-   B = m->nbMdctBlocks*m->nbChannels;
-   for (i=0;i<m->nbEBands;i++)
+   B = m->nbMdctBlocks;
+   C = m->nbChannels;
+   for (c=0;c<C;c++)
    {
-      int j;
-      float x = bank[i];
-      for (j=B*eBands[i];j<B*eBands[i+1];j++)
-         X[j] *= x;
+      for (i=0;i<m->nbEBands;i++)
+      {
+         int j;
+         float g = bank[i*C+c];
+         for (j=B*eBands[i];j<B*eBands[i+1];j++)
+            X[j*C+c] *= g;
+      }
    }
-   for (i=B*eBands[m->nbEBands];i<B*eBands[m->nbEBands+1];i++)
+   for (i=B*C*eBands[m->nbEBands];i<B*C*eBands[m->nbEBands+1];i++)
       X[i] = 0;
 }
 
diff --git a/libcelt/bands.h b/libcelt/bands.h
index 204cb794..af15e890 100644
--- a/libcelt/bands.h
+++ b/libcelt/bands.h
@@ -52,6 +52,8 @@ void compute_band_energies(const CELTMode *m, float *X, float *bands);
  */
 void normalise_bands(const CELTMode *m, float *X, float *bands);
 
+void renormalise_bands(const CELTMode *m, float *X);
+
 /** Denormalise each band of X to restore full amplitude
  * @param m Mode data 
  * @param X Spectrum (returned de-normalised)
diff --git a/libcelt/celt.c b/libcelt/celt.c
index 941ac7d2..c165762a 100644
--- a/libcelt/celt.c
+++ b/libcelt/celt.c
@@ -239,7 +239,7 @@ int celt_encode(CELTEncoder *st, short *pcm)
    float X[B*C*N];         /**< Interleaved signal MDCTs */
    float P[B*C*N];         /**< Interleaved pitch MDCTs*/
    float mask[B*C*N];      /**< Masking curve */
-   float bandE[st->mode->nbEBands];
+   float bandE[st->mode->nbEBands*C];
    float gains[st->mode->nbPBands];
    int pitch_index;
 
@@ -295,28 +295,30 @@ int celt_encode(CELTEncoder *st, short *pcm)
    for (j=0;j<B*N;j++)
       printf ("%f ", P[j]);
    printf ("\n");*/
-   if (C==2)
-   {
-      haar1(X, B*N*C, 1);
-      haar1(P, B*N*C, 1);
-   }
-   
-   /* Get a tiny bit more frequency resolution and prevent unstable energy when quantising */
-   time_dct(X, N, B, C);
-   time_dct(P, N, B, C);
 
    /* Band normalisation */
    compute_band_energies(st->mode, X, bandE);
    normalise_bands(st->mode, X, bandE);
    //for (i=0;i<st->mode->nbEBands;i++)printf("%f ", bandE[i]);printf("\n");
+   //for (i=0;i<N*B*C;i++)printf("%f ", X[i]);printf("\n");
 
    /* Normalise the pitch vector as well (discard the energies) */
    {
-      float bandEp[st->mode->nbEBands];
+      float bandEp[st->mode->nbEBands*st->mode->nbChannels];
       compute_band_energies(st->mode, P, bandEp);
       normalise_bands(st->mode, P, bandEp);
    }
 
+   if (C==2)
+   {
+      haar1(X, B*N*C, 1);
+      haar1(P, B*N*C, 1);
+   }
+   /* Get a tiny bit more frequency resolution and prevent unstable energy when quantising */
+   time_dct(X, N, B, C);
+   time_dct(P, N, B, C);
+
+
    quant_energy(st->mode, bandE, st->oldBandE, &st->enc);
 
    /* Pitch prediction */
@@ -335,14 +337,16 @@ int celt_encode(CELTEncoder *st, short *pcm)
    printf ("%f\n", sum);*/
    /* Residual quantisation */
    quant_bands(st->mode, X, P, mask, &st->enc);
-
-   /* Synthesis */
-   denormalise_bands(st->mode, X, bandE);
-
+   
    time_idct(X, N, B, C);
    if (C==2)
       haar1(X, B*N*C, 1);
 
+   renormalise_bands(st->mode, X);
+   /* Synthesis */
+   denormalise_bands(st->mode, X, bandE);
+
+
    CELT_MOVE(st->out_mem, st->out_mem+C*B*N, C*(MAX_PERIOD-B*N));
 
    compute_inv_mdcts(&st->mdct_lookup, st->window, X, st->out_mem, st->mdct_overlap, N, st->overlap, B, C);
diff --git a/libcelt/modes.c b/libcelt/modes.c
index 4e625df3..8d8ee695 100644
--- a/libcelt/modes.c
+++ b/libcelt/modes.c
@@ -163,16 +163,16 @@ const CELTMode mode3 = {
 /* Stereo mode around 120 kbps */
 const CELTMode mode4 = {
    128,         /**< overlap */
-   128,         /**< mdctSize */
-   2,           /**< nbMdctBlocks */
+   256,         /**< mdctSize */
+   1,           /**< nbMdctBlocks */
    2,           /**< channels */
    
-   NBANDS128,   /**< nbEBands */
-   PBANDS128,   /**< nbPBands */
-   PITCH_END128,/**< pitchEnd */
+   NBANDS256,   /**< nbEBands */
+   PBANDS256,   /**< nbPBands */
+   PITCH_END256,/**< pitchEnd */
    
-   qbank1,      /**< eBands */
-   pbank1,      /**< pBands*/
+   qbank3,      /**< eBands */
+   pbank3,      /**< pBands*/
    qpulses2s,   /**< nbPulses */
    
    0.7,         /**< ePredCoef */