diff --git a/src/tuimble/audio.py b/src/tuimble/audio.py
index 1262a6b..9df9e36 100644
--- a/src/tuimble/audio.py
+++ b/src/tuimble/audio.py
@@ -1,7 +1,8 @@
 """Audio capture and playback pipeline.
 
-Handles microphone input, speaker output, and Opus encoding/decoding.
-Designed to run alongside the async event loop via callback-based I/O.
+Handles microphone input (PCM -> Opus) and speaker output (PCM direct).
+pymumble already decodes incoming audio, so the playback path accepts raw
+PCM — no decoder needed.  The encoder is used only for the capture path.
 """
 
 from __future__ import annotations
@@ -36,7 +37,6 @@ class AudioPipeline:
         self._playback_queue: queue.Queue[bytes] = queue.Queue(maxsize=50)
 
         self._encoder = None
-        self._decoder = None
         self._input_stream = None
         self._output_stream = None
         self._capturing = False
@@ -49,7 +49,6 @@ class AudioPipeline:
         self._encoder = opuslib.Encoder(
             self._sample_rate, CHANNELS, opuslib.APPLICATION_VOIP
         )
-        self._decoder = opuslib.Decoder(self._sample_rate, CHANNELS)
 
         self._output_stream = sd.RawOutputStream(
             samplerate=self._sample_rate,
@@ -107,12 +106,11 @@ class AudioPipeline:
         if status:
             log.warning("playback status: %s", status)
         try:
-            data = self._playback_queue.get_nowait()
-            if self._decoder:
-                pcm = self._decoder.decode(data, self._frame_size)
-                outdata[:] = pcm[: len(outdata)]
-            else:
-                outdata[:] = b"\x00" * len(outdata)
+            pcm = self._playback_queue.get_nowait()
+            n = min(len(pcm), len(outdata))
+            outdata[:n] = pcm[:n]
+            if n < len(outdata):
+                outdata[n:] = b"\x00" * (len(outdata) - n)
         except queue.Empty:
             outdata[:] = b"\x00" * len(outdata)
 
@@ -123,9 +121,9 @@ class AudioPipeline:
         except queue.Empty:
             return None
 
-    def queue_playback(self, opus_data: bytes):
-        """Queue encoded audio data for playback."""
+    def queue_playback(self, pcm_data: bytes):
+        """Queue raw PCM data for playback (16-bit, mono, 48kHz)."""
         try:
-            self._playback_queue.put_nowait(opus_data)
+            self._playback_queue.put_nowait(pcm_data)
         except queue.Full:
             pass