diff --git a/media/libmedia/AudioTrack.cpp b/media/libmedia/AudioTrack.cpp
index ccd1b49..11b0b89 100644
--- a/media/libmedia/AudioTrack.cpp
+++ b/media/libmedia/AudioTrack.cpp
@@ -879,7 +879,8 @@
     ALOGV("createTrack_l() output %d afLatency %d", output, afLatency);
 
     // The client's AudioTrack buffer is divided into n parts for purpose of wakeup by server, where
-    //  n = 1   fast track; nBuffering is ignored
+    //  n = 1   fast track with single buffering; nBuffering is ignored
+    //  n = 2   fast track with double buffering
     //  n = 2   normal track, no sample rate conversion
     //  n = 3   normal track, with sample rate conversion
     //          (pessimistic; some non-1:1 conversion ratios don't actually need triple-buffering)
@@ -1019,9 +1020,11 @@
             ALOGV("AUDIO_OUTPUT_FLAG_FAST successful; frameCount %u", frameCount);
             mAwaitBoost = true;
             if (sharedBuffer == 0) {
-                // double-buffering is not required for fast tracks, due to tighter scheduling
-                if (mNotificationFramesAct == 0 || mNotificationFramesAct > frameCount) {
-                    mNotificationFramesAct = frameCount;
+                // Theoretically double-buffering is not required for fast tracks,
+                // due to tighter scheduling.  But in practice, to accommodate kernels with
+                // scheduling jitter, and apps with computation jitter, we use double-buffering.
+                if (mNotificationFramesAct == 0 || mNotificationFramesAct > frameCount/nBuffering) {
+                    mNotificationFramesAct = frameCount/nBuffering;
                 }
             }
         } else {