diff --git a/libcelt/bands.c b/libcelt/bands.c
index 05a375e8..d30876e8 100644
--- a/libcelt/bands.c
+++ b/libcelt/bands.c
@@ -289,8 +289,88 @@ void pitch_quant_bands(const CELTMode *m, celt_norm_t * restrict P, const celt_p
       P[i] = 0;
 }
 
+static void intensity_band(celt_norm_t * restrict X, int len)
+{
+   int j;
+   celt_word32_t E = 1e-15;
+   celt_word32_t E2 = 1e-15;
+   for (j=0;j<len;j++)
+   {
+      X[j] = X[2*j];
+      E += MULT16_16(X[j],X[j]);
+      E2 += MULT16_16(X[2*j+1],X[2*j+1]);
+   }
+#ifndef FIXED_POINT
+   E  = celt_sqrt(E+E2)/celt_sqrt(E);
+   for (j=0;j<len;j++)
+      X[j] *= E;
+#endif
+   for (j=0;j<len;j++)
+      X[len+j] = 0;
+
+}
+
+static void dup_band(celt_norm_t * restrict X, int len)
+{
+   int j;
+   for (j=len-1;j>=0;j--)
+   {
+      X[2*j] = MULT16_16_Q15(QCONST16(.70711f,15),X[j]);
+      X[2*j+1] = MULT16_16_Q15(QCONST16(.70711f,15),X[j]);
+   }
+}
+
+static void stereo_band_mix(const CELTMode *m, celt_norm_t *X, const celt_ener_t *bank, const int *stereo_mode, int bandID, int dir)
+{
+   int i = bandID;
+   const celt_int16_t *eBands = m->eBands;
+   const int C = CHANNELS(m);
+   {
+      int j;
+      celt_word16_t left, right;
+      celt_word16_t a1, a2;
+      celt_word16_t norm;
+#ifdef FIXED_POINT
+      int shift = celt_zlog2(MAX32(bank[i*C], bank[i*C+1]))-13;
+#endif
+      left = VSHR32(bank[i*C],shift);
+      right = VSHR32(bank[i*C+1],shift);
+      norm = EPSILON + celt_sqrt(EPSILON+MULT16_16(left,left)+MULT16_16(right,right));
+      a1 = DIV32_16(SHL32(EXTEND32(left),14),norm);
+      a2 = dir*DIV32_16(SHL32(EXTEND32(right),14),norm);
+      if (stereo_mode[i] && dir <0)
+      {
+         dup_band(X+C*eBands[i], eBands[i+1]-eBands[i]);
+      } else {
+         for (j=eBands[i];j<eBands[i+1];j++)
+         {
+            celt_norm_t r, l;
+            l = X[j*C];
+            r = X[j*C+1];
+            X[j*C] = MULT16_16_Q14(a1,l) + MULT16_16_Q14(a2,r);
+            X[j*C+1] = MULT16_16_Q14(a1,r) - MULT16_16_Q14(a2,l);
+         }
+      }
+      if (stereo_mode[i] && dir>0)
+      {
+         intensity_band(X+C*eBands[i], eBands[i+1]-eBands[i]);
+      }
+   }
+}
+
+void stereo_decision(const CELTMode *m, celt_norm_t * restrict X, int *stereo_mode, int len)
+{
+   int i;
+   for (i=0;i<len-5;i++)
+      stereo_mode[i] = 0;
+   for (;i<len;i++)
+      stereo_mode[i] = 1;
+}
+
+
+
 /* Quantisation of the residual */
-void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, celt_mask_t *W, const int *stereo_mode, int total_bits, ec_enc *enc)
+void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, celt_mask_t *W, const celt_ener_t *bandE, const int *stereo_mode, int total_bits, ec_enc *enc)
 {
    int i, j, bits;
    const celt_int16_t * restrict eBands = m->eBands;
@@ -336,15 +416,20 @@ void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, ce
       
       if (q > 0)
       {
-         /*int nb_rotations = q <= 2*C ? 2*C/q : 0;
-         if (nb_rotations != 0)
+         int ch=C;
+         if (C==2 && stereo_mode[i]==1)
+            ch = 1;
+         if (C==2)
          {
-            exp_rotation(P+C*eBands[i], C*(eBands[i+1]-eBands[i]), -1, C, nb_rotations);
-            exp_rotation(X+C*eBands[i], C*(eBands[i+1]-eBands[i]), -1, C, nb_rotations);
-         }*/
-         alg_quant(X+C*eBands[i], W+C*eBands[i], C*(eBands[i+1]-eBands[i]), q, P+C*eBands[i], enc);
-         /*if (nb_rotations != 0)
-            exp_rotation(X+C*eBands[i], C*(eBands[i+1]-eBands[i]), 1, C, nb_rotations);*/
+            stereo_band_mix(m, X, bandE, stereo_mode, i, 1);
+            stereo_band_mix(m, P, bandE, stereo_mode, i, 1);
+         }
+         alg_quant(X+C*eBands[i], W+C*eBands[i], ch*(eBands[i+1]-eBands[i]), q, P+C*eBands[i], enc);
+         if (C==2)
+            stereo_band_mix(m, X, bandE, stereo_mode, i, -1);
+      } else {
+         for (j=C*eBands[i];j<C*eBands[i+1];j++)
+            X[j] = P[j];
       }
       for (j=C*eBands[i];j<C*eBands[i+1];j++)
          norm[j] = MULT16_16_Q15(n,X[j]);
@@ -353,7 +438,7 @@ void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, ce
 }
 
 /* Decoding of the residual */
-void unquant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, const int *stereo_mode, int total_bits, ec_dec *dec)
+void unquant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, const celt_ener_t *bandE, const int *stereo_mode, int total_bits, ec_dec *dec)
 {
    int i, j, bits;
    const celt_int16_t * restrict eBands = m->eBands;
@@ -394,12 +479,17 @@ void unquant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P,
       
       if (q > 0)
       {
-         /*int nb_rotations = q <= 2*C ? 2*C/q : 0;
-         if (nb_rotations != 0)
-            exp_rotation(P+C*eBands[i], C*(eBands[i+1]-eBands[i]), -1, C, nb_rotations);*/
-         alg_unquant(X+C*eBands[i], C*(eBands[i+1]-eBands[i]), q, P+C*eBands[i], dec);
-         /*if (nb_rotations != 0)
-            exp_rotation(X+C*eBands[i], C*(eBands[i+1]-eBands[i]), 1, C, nb_rotations);*/
+         int ch=C;
+         if (C==2 && stereo_mode[i]==1)
+            ch = 1;
+         if (C==2)
+            stereo_band_mix(m, P, bandE, stereo_mode, i, 1);
+         alg_unquant(X+C*eBands[i], ch*(eBands[i+1]-eBands[i]), q, P+C*eBands[i], dec);
+         if (C==2)
+            stereo_band_mix(m, X, bandE, stereo_mode, i, -1);
+      } else {
+         for (j=C*eBands[i];j<C*eBands[i+1];j++)
+            X[j] = P[j];
       }
       for (j=C*eBands[i];j<C*eBands[i+1];j++)
          norm[j] = MULT16_16_Q15(n,X[j]);
@@ -407,44 +497,3 @@ void unquant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P,
    RESTORE_STACK;
 }
 
-#ifndef DISABLE_STEREO
-void stereo_decision(const CELTMode *m, celt_norm_t * restrict X, int *stereo_mode, int len)
-{
-   int i;
-   for (i=0;i<len-5;i++)
-      stereo_mode[i] = 0;
-   for (;i<len;i++)
-      stereo_mode[i] = 0;
-}
-
-
-void stereo_mix(const CELTMode *m, celt_norm_t *X, const celt_ener_t *bank, const int *stereo_mode, int dir)
-{
-   int i;
-   const celt_int16_t *eBands = m->eBands;
-   const int C = CHANNELS(m);
-   for (i=0;i<m->nbEBands;i++)
-   {
-      int j;
-      celt_word16_t left, right;
-      celt_word16_t a1, a2;
-      celt_word16_t norm;
-#ifdef FIXED_POINT
-      int shift = celt_zlog2(MAX32(bank[i*C], bank[i*C+1]))-13;
-#endif
-      left = VSHR32(bank[i*C],shift);
-      right = VSHR32(bank[i*C+1],shift);
-      norm = EPSILON + celt_sqrt(EPSILON+MULT16_16(left,left)+MULT16_16(right,right));
-      a1 = DIV32_16(SHL32(EXTEND32(left),14),norm);
-      a2 = dir*DIV32_16(SHL32(EXTEND32(right),14),norm);
-      for (j=eBands[i];j<eBands[i+1];j++)
-      {
-         celt_norm_t r, l;
-         l = X[j*C];
-         r = X[j*C+1];
-         X[j*C] = MULT16_16_Q14(a1,l) + MULT16_16_Q14(a2,r);
-         X[j*C+1] = MULT16_16_Q14(a1,r) - MULT16_16_Q14(a2,l);
-      }
-   }
-}
-#endif
diff --git a/libcelt/bands.h b/libcelt/bands.h
index b5a620fc..82dce75a 100644
--- a/libcelt/bands.h
+++ b/libcelt/bands.h
@@ -86,7 +86,7 @@ void pitch_quant_bands(const CELTMode *m, celt_norm_t * restrict P, const celt_p
  * @param total_bits Total number of bits that can be used for the frame (including the ones already spent)
  * @param enc Entropy encoder
  */
-void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, celt_mask_t *W, const int *stereo_mode, int total_bits, ec_enc *enc);
+void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, celt_mask_t *W, const celt_ener_t *bandE, const int *stereo_mode, int total_bits, ec_enc *enc);
 
 /** Decoding of the residual spectrum
  * @param m Mode data 
@@ -95,10 +95,8 @@ void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, ce
  * @param total_bits Total number of bits that can be used for the frame (including the ones already spent)
  * @param dec Entropy decoder
 */
-void unquant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, const int *stereo_mode, int total_bits, ec_dec *dec);
+void unquant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, const celt_ener_t *bandE, const int *stereo_mode, int total_bits, ec_dec *dec);
 
 void stereo_decision(const CELTMode *m, celt_norm_t * restrict X, int *stereo_mode, int len);
 
-void stereo_mix(const CELTMode *m, celt_norm_t *X, const celt_ener_t *bank, const int *stereo_mode, int dir);
-
 #endif /* BANDS_H */
diff --git a/libcelt/celt.c b/libcelt/celt.c
index 962db5f3..ed867dcf 100644
--- a/libcelt/celt.c
+++ b/libcelt/celt.c
@@ -325,26 +325,17 @@ int EXPORT celt_encode(CELTEncoder * restrict st, celt_int16_t * restrict pcm, u
    quant_energy(st->mode, bandE, st->oldBandE, 20+nbCompressedBytes*8/5, st->mode->prob, &st->enc);
 
    ALLOC(stereo_mode, st->mode->nbEBands, int);
-   if (C==2)
-   {
-      stereo_decision(st->mode, X, stereo_mode, st->mode->nbEBands);
-      stereo_mix(st->mode, X, bandE, stereo_mode, 1);
-      stereo_mix(st->mode, P, bandE, stereo_mode, 1);
-   }
+   stereo_decision(st->mode, X, stereo_mode, st->mode->nbEBands);
 
    pitch_quant_bands(st->mode, P, gains);
 
    /*for (i=0;i<B*N;i++) printf("%f ",P[i]);printf("\n");*/
-   /* Compute residual that we're going to encode */
-   for (i=0;i<C*st->mode->eBands[st->mode->nbEBands];i++)
-      X[i] -= P[i];
 
    /* Residual quantisation */
-   quant_bands(st->mode, X, P, NULL, stereo_mode, nbCompressedBytes*8, &st->enc);
+   quant_bands(st->mode, X, P, NULL, bandE, stereo_mode, nbCompressedBytes*8, &st->enc);
    
    if (C==2)
    {
-      stereo_mix(st->mode, X, bandE, stereo_mode, -1);
       renormalise_bands(st->mode, X);
    }
    /* Synthesis */
@@ -606,20 +597,15 @@ int EXPORT celt_decode(CELTDecoder * restrict st, unsigned char *data, int len,
    }
 
    ALLOC(stereo_mode, st->mode->nbEBands, int);
-   if (C==2)
-   {
-      stereo_decision(st->mode, X, stereo_mode, st->mode->nbEBands);
-      stereo_mix(st->mode, P, bandE, stereo_mode, 1);
-   }
+   stereo_decision(st->mode, X, stereo_mode, st->mode->nbEBands);
    /* Apply pitch gains */
    pitch_quant_bands(st->mode, P, gains);
 
    /* Decode fixed codebook and merge with pitch */
-   unquant_bands(st->mode, X, P, stereo_mode, len*8, &dec);
+   unquant_bands(st->mode, X, P, bandE, stereo_mode, len*8, &dec);
 
    if (C==2)
    {
-      stereo_mix(st->mode, X, bandE, stereo_mode, -1);
       renormalise_bands(st->mode, X);
    }
    /* Synthesis */
diff --git a/libcelt/dump_modes.c b/libcelt/dump_modes.c
index 9f00dcc1..de4e062c 100644
--- a/libcelt/dump_modes.c
+++ b/libcelt/dump_modes.c
@@ -156,6 +156,7 @@ void dump_modes(FILE *file, CELTMode **modes, int nb_modes)
       fprintf(file, "%d,\t/* nbAllocVectors */\n", mode->nbAllocVectors);
       fprintf(file, "allocVectors%d_%d,\t/* allocVectors */\n", mode->Fs, mode->mdctSize);
       fprintf(file, "allocCache%d_%d_%d,\t/* bits */\n", mode->Fs, mode->mdctSize, mode->nbChannels);
+      fprintf(file, "0,\t/* bits_stereo */\n");
       fprintf(file, "{%d, 0, 0},\t/* mdct */\n", 2*mode->mdctSize);
       fprintf(file, "0,\t/* fft */\n");
       fprintf(file, "window%d,\t/* window */\n", mode->overlap);
diff --git a/libcelt/modes.c b/libcelt/modes.c
index 89ba428f..a9a94e78 100644
--- a/libcelt/modes.c
+++ b/libcelt/modes.c
@@ -364,8 +364,9 @@ CELTMode EXPORT *celt_mode_create(celt_int32_t Fs, int channels, int frame_size,
 #endif
    mode->window = window;
 
-   mode->bits = (const celt_int16_t **)compute_alloc_cache(mode, mode->nbChannels);
+   mode->bits = (const celt_int16_t **)compute_alloc_cache(mode, 1);
 
+   mode->bits_stereo = NULL;
 #ifndef SHORTCUTS
    psydecay_init(&mode->psy, MAX_PERIOD/2, mode->Fs);
 #endif
@@ -379,6 +380,9 @@ CELTMode EXPORT *celt_mode_create(celt_int32_t Fs, int channels, int frame_size,
    mode->prob = quant_prob_alloc(mode);
    compute_energy_allocation_table(mode);
    
+   if (mode->nbChannels>=2)
+      mode->bits_stereo = (const celt_int16_t **)compute_alloc_cache(mode, mode->nbChannels);
+
    if (error)
       *error = CELT_OK;
    return mode;
diff --git a/libcelt/modes.h b/libcelt/modes.h
index dd926dc2..21ca35c7 100644
--- a/libcelt/modes.h
+++ b/libcelt/modes.h
@@ -89,6 +89,8 @@ struct CELTMode {
    
    const celt_int16_t * const *bits; /**< Cache for pulses->bits mapping in each band */
 
+   const celt_int16_t * const *bits_stereo; /**< Cache for pulses->bits mapping in each band */
+
    /* Stuff that could go in the {en,de}coder, but we save space this way */
    mdct_lookup mdct;
    kiss_fftr_cfg fft;
diff --git a/libcelt/rate.c b/libcelt/rate.c
index 09a62321..dbf2ca42 100644
--- a/libcelt/rate.c
+++ b/libcelt/rate.c
@@ -105,7 +105,6 @@ celt_int16_t **compute_alloc_cache(CELTMode *m, int C)
 
    bits = celt_alloc(m->nbEBands*sizeof(celt_int16_t*));
    
-   C = m->nbChannels;
    prevN = -1;
    for (i=0;i<m->nbEBands;i++)
    {
@@ -270,6 +269,8 @@ int compute_allocation(const CELTMode *m, int *offsets, const int *stereo_mode,
       for (i=0;i<len;i++)
       {
          if (stereo_mode[i]==0)
+            cache[i] = m->bits_stereo[i];
+         else
             cache[i] = m->bits[i];
       }
    } else {
diff --git a/libcelt/vq.c b/libcelt/vq.c
index 76b0ce0d..9426ff4b 100644
--- a/libcelt/vq.c
+++ b/libcelt/vq.c
@@ -118,6 +118,7 @@ void alg_quant(celt_norm_t *X, celt_mask_t *W, int N, int K, const celt_norm_t *
 
    sum = 0;
    j=0; do {
+      X[j] -= P[j];
       if (X[j]>0)
          signx[j]=1;
       else
@@ -307,14 +308,6 @@ void intra_prediction(const CELTMode *m, celt_norm_t * restrict x, celt_mask_t *
    pred_gain = s*MULT16_16_Q15(pred_gain,celt_rcp(SHL32(celt_sqrt(E),9)));
    for (j=0;j<C*N;j++)
       P[j] = PSHR32(MULT16_16(pred_gain, P[j]),8);
-   if (K>0)
-   {
-      for (j=0;j<C*N;j++)
-         x[j] -= P[j];
-   } else {
-      for (j=0;j<C*N;j++)
-         x[j] = P[j];
-   }
 }
 
 void intra_unquant(const CELTMode *m, celt_norm_t *x, int N, int K, celt_norm_t *Y, celt_norm_t * restrict P, int N0, int Nmax, ec_dec *dec)
@@ -353,11 +346,6 @@ void intra_unquant(const CELTMode *m, celt_norm_t *x, int N, int K, celt_norm_t
    pred_gain = s*MULT16_16_Q15(pred_gain,celt_rcp(SHL32(celt_sqrt(E),9)));
    for (j=0;j<C*N;j++)
       P[j] = PSHR32(MULT16_16(pred_gain, P[j]),8);
-   if (K==0)
-   {
-      for (j=0;j<C*N;j++)
-         x[j] = P[j];
-   }
 }
 
 void intra_fold(const CELTMode *m, celt_norm_t *x, int N, celt_norm_t *Y, celt_norm_t * restrict P, int N0, int Nmax)
@@ -388,7 +376,5 @@ void intra_fold(const CELTMode *m, celt_norm_t *x, int N, celt_norm_t *Y, celt_n
    g = celt_rcp(SHL32(celt_sqrt(E),9));
    for (j=0;j<C*N;j++)
       P[j] = PSHR32(MULT16_16(g, P[j]),8);
-   for (j=0;j<C*N;j++)
-      x[j] = P[j];
 }