diff --git a/Source/Core/Core/HW/DSPHLE/UCodes/Zelda.cpp b/Source/Core/Core/HW/DSPHLE/UCodes/Zelda.cpp
index c540dc2461..f11582f776 100644
--- a/Source/Core/Core/HW/DSPHLE/UCodes/Zelda.cpp
+++ b/Source/Core/Core/HW/DSPHLE/UCodes/Zelda.cpp
@@ -1332,24 +1332,23 @@ void ZeldaAudioRenderer::AddVoice(u16 voice_id)
 
     struct
     {
-      MixingBuffer* buffer;
+      MixingBuffer& buffer;
       s16 volume;
       s16 volume_delta;
     } buffers[8] = {
-        {&m_buf_front_left, quadrant_volumes[0], volume_deltas[0]},
-        {&m_buf_back_left, quadrant_volumes[1], volume_deltas[1]},
-        {&m_buf_front_right, quadrant_volumes[2], volume_deltas[2]},
-        {&m_buf_back_right, quadrant_volumes[3], volume_deltas[3]},
+        {m_buf_front_left, quadrant_volumes[0], volume_deltas[0]},
+        {m_buf_back_left, quadrant_volumes[1], volume_deltas[1]},
+        {m_buf_front_right, quadrant_volumes[2], volume_deltas[2]},
+        {m_buf_back_right, quadrant_volumes[3], volume_deltas[3]},
 
-        {&m_buf_front_left_reverb, reverb_volumes[0], reverb_volume_deltas[0]},
-        {&m_buf_back_left_reverb, reverb_volumes[1], reverb_volume_deltas[1]},
-        {&m_buf_front_right_reverb, reverb_volumes[2], reverb_volume_deltas[2]},
-        {&m_buf_back_right_reverb, reverb_volumes[3], reverb_volume_deltas[3]},
+        {m_buf_front_left_reverb, reverb_volumes[0], reverb_volume_deltas[0]},
+        {m_buf_back_left_reverb, reverb_volumes[1], reverb_volume_deltas[1]},
+        {m_buf_front_right_reverb, reverb_volumes[2], reverb_volume_deltas[2]},
+        {m_buf_back_right_reverb, reverb_volumes[3], reverb_volume_deltas[3]},
     };
     for (const auto& buffer : buffers)
     {
-      AddBuffersWithVolumeRamp(buffer.buffer, input_samples, buffer.volume << 16,
-                               (buffer.volume_delta << 16) / (s32)buffer.buffer->size());
+      AddBuffersWithVolumeRamp(buffer.buffer, input_samples, buffer.volume, buffer.volume_delta);
     }
 
     vpb.dolby_volume_current = vpb.dolby_volume_target;
@@ -1384,13 +1383,11 @@ void ZeldaAudioRenderer::AddVoice(u16 voice_id)
       else
         volume_delta = vpb.channels[i].target_volume - vpb.channels[i].current_volume;
 
-      s32 volume_step = (volume_delta << 16) / (s32)input_samples.size();  // In 1.31 format.
-
       // TODO: The last value of each channel structure is used to
       // determine whether a channel should be skipped or not. Not
       // implemented yet.
 
-      if (!vpb.channels[i].current_volume && !volume_step)
+      if (!vpb.channels[i].current_volume && !volume_delta)
         continue;
 
       MixingBuffer* dst_buffer = BufferForID(vpb.channels[i].id);
@@ -1402,9 +1399,9 @@ void ZeldaAudioRenderer::AddVoice(u16 voice_id)
         continue;
       }
 
-      s32 new_volume = AddBuffersWithVolumeRamp(dst_buffer, input_samples,
-                                                vpb.channels[i].current_volume << 16, volume_step);
-      vpb.channels[i].current_volume = new_volume >> 16;
+      s16 new_volume = AddBuffersWithVolumeRamp(*dst_buffer, input_samples,
+                                                vpb.channels[i].current_volume, volume_delta);
+      vpb.channels[i].current_volume = new_volume;
     }
   }
 
diff --git a/Source/Core/Core/HW/DSPHLE/UCodes/Zelda.h b/Source/Core/Core/HW/DSPHLE/UCodes/Zelda.h
index 7bc4504f4e..39d06d9077 100644
--- a/Source/Core/Core/HW/DSPHLE/UCodes/Zelda.h
+++ b/Source/Core/Core/HW/DSPHLE/UCodes/Zelda.h
@@ -51,6 +51,8 @@ private:
   // See Zelda.cpp for the list of possible flags.
   u32 m_flags;
 
+  typedef std::array<s16, 0x50> MixingBuffer;
+
   // Utility functions for audio operations.
 
   // Apply volume to a buffer. The volume is a fixed point integer, usually
@@ -77,25 +79,33 @@ private:
     ApplyVolumeInPlace<N, 4>(buf, vol);
   }
 
-  // Mixes two buffers together while applying a volume to one of them. The
-  // volume ramps up/down in N steps using the provided step delta value.
-  //
-  // Note: On a real GC, the stepping happens in 32 steps instead. But hey,
-  // we can do better here with very low risk. Why not? :)
-  template <size_t N>
-  s32 AddBuffersWithVolumeRamp(std::array<s16, N>* dst, const std::array<s16, N>& src, s32 vol,
-                               s32 step)
+  // Mixes two buffers together while applying a volume to one of them.
+  // We try to match LLE, hence the slightly odd ramping.
+  s16 AddBuffersWithVolumeRamp(MixingBuffer& dst, const MixingBuffer& src, s16 start_volume,
+                               s16 delta)
   {
-    if (!vol && !step)
-      return vol;
+    if (!start_volume && !delta)
+      return start_volume;
 
-    for (size_t i = 0; i < N; ++i)
+    // The delta is applied in 32 steps over the first 64 samples.
+    s32 vol = start_volume << 16;
+    s32 step = delta << (16 - 5);
+    for (size_t i = 0; i < 0x40;)
     {
-      (*dst)[i] += ((vol >> 16) * src[i]) >> 16;
+      dst[i] += ((vol >> 16) * src[i]) >> 16;
+      ++i;
+      dst[i] += ((vol >> 16) * src[i]) >> 16;
+      ++i;
       vol += step;
     }
 
-    return vol;
+    // The last 16 samples are mixed at the target volume.
+    for (size_t i = 0x40; i < 0x50; ++i)
+    {
+      dst[i] += ((vol >> 16) * src[i]) >> 16;
+    }
+
+    return vol >> 16;
   }
 
   // Does not use std::array because it needs to be able to process partial
@@ -120,7 +130,6 @@ private:
   u16 m_output_volume = 0;
 
   // Mixing buffers.
-  typedef std::array<s16, 0x50> MixingBuffer;
   MixingBuffer m_buf_front_left{};
   MixingBuffer m_buf_front_right{};
   MixingBuffer m_buf_back_left{};