diff --git a/sendspin/audio.py b/sendspin/audio.py
index 5bf1a63..099dc28 100644
--- a/sendspin/audio.py
+++ b/sendspin/audio.py
@@ -253,6 +253,14 @@ def set_format(self, audio_format: AudioFormat, device: AudioDevice) -> None:
             device: Audio device to use.
         """
         pcm_format = audio_format.pcm_format
+        if self._stream is not None and self._format == pcm_format:
+            logger.debug(
+                "Audio format unchanged (%dHz/%d-bit/%dch); keeping stream open",
+                pcm_format.sample_rate,
+                pcm_format.bit_depth,
+                pcm_format.channels,
+            )
+            return
         self._format = pcm_format
         self._close_stream()
 
@@ -340,17 +348,16 @@ def stop(self) -> None:
         self._stream_executor.shutdown(wait=True)
 
     def close_stream(self) -> None:
-        """Drop queued audio and fully close the stream to release the device.
+        """Drop queued audio but keep the stream open on silence.
 
-        Unlike clear(), which only stops the stream (leaving the device FD open),
-        this fully closes the PortAudio stream. Call when the server signals
-        end-of-stream; the stream will be recreated by set_format() when the
-        next track begins.
+        The callback renders silence while starved, so the audio device
+        stays open and DACs hold lock across pause/resume and across
+        stream end/start pairs. The device is released by stop()
+        (shutdown) and by set_format() when the format actually changes.
         """
         if self._closed:
             return
         self.clear()
-        self._close_stream()
 
     def clear(self) -> None:
         """Drop all queued audio chunks."""
@@ -359,13 +366,11 @@ def clear(self) -> None:
         # Clear deferred operation flag
         self._clear_requested = False
 
-        # Stop the audio stream (but don't close it) to release ALSA device
-        # This allows the device to transition to 'closed' state when paused.
-        # The stream restarts when new chunks arrive in submit().
-        self._stream_started = False
-        stream = self._stream
-        if stream is not None:
-            self._stream_executor.submit(self._call_stream, stream.stop)
+        # Keep the audio stream running: with the queue drained the callback
+        # renders silence, so the device stays open and DACs that click a
+        # muting relay on lock loss (e.g. XMOS USB) stay silent across
+        # pause/resume/seek. Chunks submitted after the clear play on the
+        # still-running stream; submit() restarts it only if never started.
 
         # Drain all queued chunks
         while True:
diff --git a/tests/test_audio.py b/tests/test_audio.py
index 3290cb5..2d871a3 100644
--- a/tests/test_audio.py
+++ b/tests/test_audio.py
@@ -61,3 +61,91 @@ def now() -> int:
 
     assert player._drop_every_n_frames == 0  # noqa: SLF001
     assert player._insert_every_n_frames == 0  # noqa: SLF001
+
+
+class _RecordingStream:
+    def __init__(self) -> None:
+        self.stops = 0
+
+    def stop(self) -> None:
+        self.stops += 1
+
+
+def test_clear_keeps_stream_open_so_device_never_closes() -> None:
+    """clear() must drain the queue without stopping the stream.
+
+    The callback renders silence while starved, so keeping the stream
+    running holds the audio device open and DACs with lock-loss muting
+    relays (e.g. XMOS USB) stay silent across pause/resume/seek.
+    """
+    player = AudioPlayer(lambda ts: ts, lambda ts: ts, now_us=lambda: 0)
+    stream = _RecordingStream()
+    player._stream = stream  # noqa: SLF001
+    player._stream_started = True  # noqa: SLF001
+    player._queue.put(  # noqa: SLF001
+        _QueuedChunk(server_timestamp_us=0, audio_data=b"\x01\x00\x02\x00")
+    )
+
+    player.clear()
+
+    assert player._queue.empty()  # noqa: SLF001
+    assert player._stream_started  # noqa: SLF001
+    assert stream.stops == 0
+
+
+class _RecordingStreamFull:
+    def __init__(self) -> None:
+        self.stops = 0
+        self.starts = 0
+        self.closes = 0
+
+    def stop(self) -> None:
+        self.stops += 1
+
+    def start(self) -> None:
+        self.starts += 1
+
+    def close(self) -> None:
+        self.closes += 1
+
+
+def _make_player() -> AudioPlayer:
+    return AudioPlayer(lambda ts: ts, lambda ts: ts, now_us=lambda: 0)
+
+
+def test_close_stream_keeps_device_open_on_silence() -> None:
+    """End-of-stream must not release the audio device.
+
+    The server sends stream/end on pause, so fully closing here is what
+    drops DAC lock and clicks the relay. Silence keeps it held.
+    """
+    player = _make_player()
+    stream = _RecordingStreamFull()
+    player._stream = stream  # noqa: SLF001
+    player._stream_started = True  # noqa: SLF001
+    player._queue.put(  # noqa: SLF001
+        _QueuedChunk(server_timestamp_us=0, audio_data=b"\x01\x00\x02\x00")
+    )
+
+    player.close_stream()
+
+    assert player._queue.empty()  # noqa: SLF001
+    assert player._stream_started  # noqa: SLF001
+    assert player._stream is stream  # noqa: SLF001
+    assert (stream.stops, stream.closes) == (0, 0)
+
+
+def test_set_format_same_format_keeps_existing_stream() -> None:
+    """Resume after pause re-sends the identical format: no recreate."""
+    from aiosendspin.client import PCMFormat
+
+    player = _make_player()
+    stream = _RecordingStreamFull()
+    fmt = PCMFormat(sample_rate=48000, channels=2, bit_depth=16)
+    player._format = fmt  # noqa: SLF001
+    player._stream = stream  # noqa: SLF001
+
+    player.set_format(SimpleNamespace(pcm_format=fmt), device=None)
+
+    assert player._stream is stream  # noqa: SLF001
+    assert (stream.stops, stream.closes) == (0, 0)