diff --git a/media/gpu/windows/mf_audio_encoder.cc b/media/gpu/windows/mf_audio_encoder.cc
--- a/media/gpu/windows/mf_audio_encoder.cc
+++ b/media/gpu/windows/mf_audio_encoder.cc
@@ -327,11 +327,15 @@ HRESULT GetSampleBuffer(const DWORD required_size,
     DWORD buffer_capacity;
     RETURN_IF_FAILED(buffer->GetMaxLength(&buffer_capacity));

-    if (buffer_capacity < required_size)
+    if (buffer_capacity < required_size) {
       need_buffer_allocation = true;
+      // Replace, don't append: drop the too-small buffer(s) so the sample
+      // contains exactly one buffer that the MFT writes into and we read from.
+      RETURN_IF_FAILED(sample->RemoveAllBuffers());
+    }
   }

   if (need_buffer_allocation) {
     RETURN_IF_FAILED(
         MFCreateAlignedMemoryBuffer(required_size, buffer_alignment, &buffer));
     RETURN_IF_FAILED(sample->AddBuffer(buffer.Get()));
   }