diff --git a/plugins/platforms/discord/adapter.py b/plugins/platforms/discord/adapter.py
index 88d87bc8b35c..99e0f9a2f57b 100644
--- a/plugins/platforms/discord/adapter.py
+++ b/plugins/platforms/discord/adapter.py
@@ -838,34 +838,37 @@ def flush_pending(self) -> list:
     @staticmethod
     def pcm_to_wav(pcm_data: bytes, output_path: str,
                    src_rate: int = 48000, src_channels: int = 2):
-        """Convert raw PCM to 16kHz mono WAV via ffmpeg."""
-        with tempfile.NamedTemporaryFile(suffix=".pcm", delete=False) as f:
-            f.write(pcm_data)
-            pcm_path = f.name
-        try:
-            from hermes_cli._subprocess_compat import windows_hide_flags
+        """Convert raw PCM to 16kHz mono WAV via ffmpeg.
 
-            subprocess.run(
-                [
-                    resolve_ffmpeg_executable(), "-y", "-loglevel", "error",
-                    "-f", "s16le",
-                    "-ar", str(src_rate),
-                    "-ac", str(src_channels),
-                    "-i", pcm_path,
-                    "-ar", "16000",
-                    "-ac", "1",
-                    output_path,
-                ],
-                check=True,
-                timeout=10,
-                stdin=subprocess.DEVNULL,
-                creationflags=windows_hide_flags(),
-            )
-        finally:
-            try:
-                os.unlink(pcm_path)
-            except OSError:
-                pass
+        The PCM is fed straight to ffmpeg's stdin, which avoids staging it in a
+        temp file on every utterance. The WAV is still written to *output_path*
+        rather than captured from stdout: ffmpeg cannot seek on a pipe, so a
+        piped WAV carries placeholder (0xFFFFFFFF) RIFF/data sizes that make
+        strict readers misreport the length.
+        """
+        from hermes_cli._subprocess_compat import windows_hide_flags
+
+        subprocess.run(
+            [
+                resolve_ffmpeg_executable(), "-y", "-loglevel", "error",
+                "-f", "s16le",
+                "-ar", str(src_rate),
+                "-ac", str(src_channels),
+                "-i", "pipe:0",
+                "-ar", "16000",
+                "-ac", "1",
+                output_path,
+            ],
+            input=pcm_data,
+            check=True,
+            timeout=10,
+            # Capture ffmpeg's -loglevel error output so a failure's
+            # CalledProcessError carries the actual message (parity with
+            # tools/transcription_tools' ffmpeg call sites) instead of
+            # "returned non-zero exit status N" with stderr detached.
+            stderr=subprocess.PIPE,
+            creationflags=windows_hide_flags(),
+        )
 
 
 def _read_dm_role_auth_guild() -> Optional[int]:
@@ -4481,7 +4484,18 @@ async def _process_voice_input(self, guild_id: int, user_id: int, pcm_data: byte
                     transcript=transcript,
                 )
         except Exception as e:
-            logger.warning("Voice input processing failed: %s", e, exc_info=True)
+            # CalledProcessError from pcm_to_wav carries ffmpeg's captured
+            # stderr — surface it, or the log only says "exit status N".
+            _ff_err = getattr(e, "stderr", None)
+            if _ff_err:
+                if isinstance(_ff_err, bytes):
+                    _ff_err = _ff_err.decode("utf-8", "replace")
+                logger.warning(
+                    "Voice input processing failed: %s (ffmpeg: %s)",
+                    e, _ff_err.strip(), exc_info=True,
+                )
+            else:
+                logger.warning("Voice input processing failed: %s", e, exc_info=True)
         finally:
             try:
                 os.unlink(wav_path)
diff --git a/tests/gateway/test_voice_command.py b/tests/gateway/test_voice_command.py
index b2d4ab6832a8..4a988a6e036a 100644
--- a/tests/gateway/test_voice_command.py
+++ b/tests/gateway/test_voice_command.py
@@ -1947,3 +1947,54 @@ def _fake_play(path):
         )
         # And the temp file is cleaned up afterwards.
         assert not os.path.exists(played[0]), "temp WAV was not unlinked"
+
+
+class TestPcmToWav:
+    """pcm_to_wav streams PCM through ffmpeg's stdin, not a temp file."""
+
+    def test_pcm_is_piped_to_stdin_not_staged_on_disk(self, tmp_path):
+        from plugins.platforms.discord.adapter import VoiceReceiver
+
+        out = tmp_path / "out.wav"
+        with patch("plugins.platforms.discord.adapter.subprocess.run") as run:
+            VoiceReceiver.pcm_to_wav(b"\x00\x01" * 16, str(out))
+
+        args, kwargs = run.call_args
+        cmd = args[0]
+        assert kwargs["input"] == b"\x00\x01" * 16, "PCM must be fed via stdin"
+        assert "pipe:0" in cmd, "ffmpeg must read the PCM from stdin"
+        assert cmd[-1] == str(out), (
+            "the WAV must be written to the real path; ffmpeg cannot seek on a "
+            "pipe, so a piped WAV gets placeholder RIFF/data sizes"
+        )
+        assert not any(str(a).endswith(".pcm") for a in cmd), (
+            "no temp .pcm file should be staged"
+        )
+
+    @pytest.mark.skipif(
+        __import__("shutil").which("ffmpeg") is None, reason="ffmpeg not installed",
+    )
+    def test_output_wav_header_reports_true_length(self, tmp_path):
+        """A piped-stdout WAV reports 0xFFFFFFFF sizes; the written file must not."""
+        import math
+        import struct
+        import wave
+
+        from plugins.platforms.discord.adapter import VoiceReceiver
+
+        frames = 48000  # 1s @ 48kHz stereo
+        pcm = b"".join(
+            struct.pack("<hh", v, v)
+            for v in (
+                int(20000 * math.sin(2 * math.pi * 440 * i / 48000))
+                for i in range(frames)
+            )
+        )
+        out = tmp_path / "out.wav"
+        VoiceReceiver.pcm_to_wav(pcm, str(out))
+
+        with wave.open(str(out)) as w:
+            assert w.getnchannels() == 1
+            assert w.getframerate() == 16000
+            # 48kHz -> 16kHz is a 3x decimation of a 1s clip.
+            assert w.getnframes() == 16000