diff --git a/src/dusk/audio/audiostream.c b/src/dusk/audio/audiostream.c
index d1bbc858..c625cd21 100644
--- a/src/dusk/audio/audiostream.c
+++ b/src/dusk/audio/audiostream.c
@@ -91,13 +91,12 @@ errorret_t audioStreamUpdate(audiostream_t *stream) {
   // start of the buffer (loopTo is not yet honored) since replaying the
   // exact same stream->data/dataSize needs no platform-specific changes;
   // an arbitrary loopTo offset would need slicing that buffer instead.
+  // Checked in this order (finished-check before needs-buffering) so that a
+  // loop restart falls straight through into re-buffering within this same
+  // call, instead of leaving BUFFERED cleared for the caller to notice and
+  // act on next frame - that extra frame of latency was audible as a gap
+  // at every loop boundary.
   if(
-    (stream->state & AUDIO_STREAM_STATE_PLAYING) &&
-    !(stream->state & AUDIO_STREAM_STATE_BUFFERED)
-  ) {
-    errorChain(audioStreamPlatformBuffer(stream));
-    stream->state |= AUDIO_STREAM_STATE_BUFFERED;
-  } else if(
     (stream->state & AUDIO_STREAM_STATE_PLAYING) &&
     (stream->state & AUDIO_STREAM_STATE_BUFFERED) &&
     audioStreamPlatformIsFinished(stream)
@@ -117,6 +116,14 @@ errorret_t audioStreamUpdate(audiostream_t *stream) {
     }
   }
 
+  if(
+    (stream->state & AUDIO_STREAM_STATE_PLAYING) &&
+    !(stream->state & AUDIO_STREAM_STATE_BUFFERED)
+  ) {
+    errorChain(audioStreamPlatformBuffer(stream));
+    stream->state |= AUDIO_STREAM_STATE_BUFFERED;
+  }
+
   errorOk();
 }
 
diff --git a/src/dusklinux/audio/audiostreamlinux.c b/src/dusklinux/audio/audiostreamlinux.c
index df0881aa..9f4bcb7c 100644
--- a/src/dusklinux/audio/audiostreamlinux.c
+++ b/src/dusklinux/audio/audiostreamlinux.c
@@ -10,6 +10,11 @@
 #include "assert/assert.h"
 #include "util/memory.h"
 
+// How many frames of lead time to keep queued ahead of playback. Matches
+// SDL_AudioSpec.samples below - the device's own internal buffer size - so
+// this is "never let the queue run drier than one SDL-internal buffer."
+#define AUDIO_LINUX_LEAD_FRAMES 4096
+
 errorret_t audioStreamLinuxInit(audiostream_t *stream) {
   assertNotNull(stream, "Stream cannot be NULL.");
 
@@ -18,7 +23,7 @@ errorret_t audioStreamLinuxInit(audiostream_t *stream) {
   desired.freq = (int) stream->pcm.sampleRate;
   desired.format = AUDIO_S16SYS;
   desired.channels = stream->pcm.channels;
-  desired.samples = 4096;
+  desired.samples = AUDIO_LINUX_LEAD_FRAMES;
 
   stream->platform.device = SDL_OpenAudioDevice(NULL, 0, &desired, NULL, 0);
   if(stream->platform.device == 0) {
@@ -63,5 +68,16 @@ errorret_t audioStreamLinuxBuffer(audiostream_t *stream) {
 bool_t audioStreamLinuxIsFinished(audiostream_t *stream) {
   assertNotNull(stream, "Stream cannot be NULL.");
 
-  return SDL_GetQueuedAudioSize(stream->platform.device) == 0;
+  // Reports "finished" (ready to be re-buffered) with AUDIO_LINUX_LEAD_FRAMES
+  // of margin still queued, rather than waiting for the queue to actually
+  // run dry. SDL_QueueAudio only ever appends to a FIFO, so queuing the next
+  // loop this early never causes overlap - it just avoids ever going silent
+  // while our once-per-frame Update() notices and catches up. Waiting for
+  // truly empty (as this used to) guarantees a gap by definition: silence
+  // has already started by the time "empty" can be observed.
+  const Uint32 leadBytes = (Uint32) (
+    AUDIO_LINUX_LEAD_FRAMES * stream->pcm.channels * sizeof(int16_t)
+  );
+
+  return SDL_GetQueuedAudioSize(stream->platform.device) <= leadBytes;
 }