diff --git a/plugins/voice.py b/plugins/voice.py
new file mode 100644
index 0000000..f6f4d0e
--- /dev/null
+++ b/plugins/voice.py
@@ -0,0 +1,309 @@
+"""Plugin: voice STT/TTS for Mumble channels.
+
+Listens for voice audio via pymumble's sound callback, buffers PCM per
+user, transcribes via Whisper STT on silence, and provides TTS playback
+via Piper.  Commands: !listen, !say.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import io
+import json
+import logging
+import threading
+import time
+import wave
+from urllib.parse import quote_plus
+import urllib.request
+
+from derp.http import urlopen as _urlopen
+from derp.plugin import command
+
+log = logging.getLogger(__name__)
+
+# -- Constants ---------------------------------------------------------------
+
+_SAMPLE_RATE = 48000
+_CHANNELS = 1
+_SAMPLE_WIDTH = 2  # s16le = 2 bytes per sample
+
+_SILENCE_GAP = 1.5      # seconds of silence before flushing
+_MIN_DURATION = 0.5      # discard utterances shorter than this
+_MAX_DURATION = 30.0     # cap buffer at this many seconds
+_MIN_BYTES = int(_MIN_DURATION * _SAMPLE_RATE * _SAMPLE_WIDTH)
+_MAX_BYTES = int(_MAX_DURATION * _SAMPLE_RATE * _SAMPLE_WIDTH)
+_FLUSH_INTERVAL = 0.5    # flush monitor poll interval
+_MAX_SAY_LEN = 500       # max characters for !say
+
+_WHISPER_URL = "http://192.168.122.1:8080/inference"
+_PIPER_URL = "http://192.168.122.1:5000/"
+
+# -- Per-bot state -----------------------------------------------------------
+
+
+def _ps(bot):
+    """Per-bot plugin runtime state."""
+    cfg = getattr(bot, "config", {}).get("voice", {})
+    return bot._pstate.setdefault("voice", {
+        "listen": False,
+        "buffers": {},          # {username: bytearray}
+        "last_ts": {},          # {username: float monotonic}
+        "flush_task": None,
+        "lock": threading.Lock(),
+        "silence_gap": cfg.get("silence_gap", _SILENCE_GAP),
+        "whisper_url": cfg.get("whisper_url", _WHISPER_URL),
+        "piper_url": cfg.get("piper_url", _PIPER_URL),
+        "_listener_registered": False,
+    })
+
+
+# -- Helpers -----------------------------------------------------------------
+
+
+def _is_mumble(bot) -> bool:
+    """Check if bot supports voice streaming."""
+    return hasattr(bot, "stream_audio")
+
+
+def _pcm_to_wav(pcm: bytes) -> bytes:
+    """Wrap raw s16le 48kHz mono PCM in a WAV container."""
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as wf:
+        wf.setnchannels(_CHANNELS)
+        wf.setsampwidth(_SAMPLE_WIDTH)
+        wf.setframerate(_SAMPLE_RATE)
+        wf.writeframes(pcm)
+    return buf.getvalue()
+
+
+# -- STT: Sound listener (pymumble thread) ----------------------------------
+
+
+def _on_voice(bot, user, sound_chunk):
+    """Buffer incoming voice PCM per user.  Runs on pymumble thread."""
+    ps = _ps(bot)
+    if not ps["listen"]:
+        return
+    name = user["name"] if isinstance(user, dict) else None
+    if not name or name == bot.nick:
+        return
+    pcm = sound_chunk.pcm
+    if not pcm:
+        return
+    with ps["lock"]:
+        if name not in ps["buffers"]:
+            ps["buffers"][name] = bytearray()
+        buf = ps["buffers"][name]
+        buf.extend(pcm)
+        if len(buf) > _MAX_BYTES:
+            ps["buffers"][name] = bytearray(buf[-_MAX_BYTES:])
+        ps["last_ts"][name] = time.monotonic()
+
+
+# -- STT: Whisper transcription ---------------------------------------------
+
+
+def _transcribe(ps, pcm: bytes) -> str:
+    """POST PCM (as WAV) to Whisper and return transcribed text.  Blocking."""
+    wav_data = _pcm_to_wav(pcm)
+    boundary = "----derp_voice_boundary"
+    body = (
+        f"--{boundary}\r\n"
+        f'Content-Disposition: form-data; name="file"; filename="audio.wav"\r\n'
+        f"Content-Type: audio/wav\r\n\r\n"
+    ).encode() + wav_data + (
+        f"\r\n--{boundary}\r\n"
+        f'Content-Disposition: form-data; name="response_format"\r\n\r\n'
+        f"json\r\n--{boundary}--\r\n"
+    ).encode()
+    req = urllib.request.Request(ps["whisper_url"], data=body, method="POST")
+    req.add_header("Content-Type", f"multipart/form-data; boundary={boundary}")
+    resp = _urlopen(req, timeout=30, proxy=False)
+    data = json.loads(resp.read())
+    resp.close()
+    return data.get("text", "").strip()
+
+
+# -- STT: Flush monitor (asyncio background task) ---------------------------
+
+
+async def _flush_monitor(bot):
+    """Poll for silence gaps and transcribe completed utterances."""
+    ps = _ps(bot)
+    loop = asyncio.get_running_loop()
+    try:
+        while ps["listen"]:
+            await asyncio.sleep(_FLUSH_INTERVAL)
+            now = time.monotonic()
+            to_flush: list[tuple[str, bytes]] = []
+
+            with ps["lock"]:
+                for name in list(ps["last_ts"]):
+                    elapsed = now - ps["last_ts"][name]
+                    if elapsed >= ps["silence_gap"] and name in ps["buffers"]:
+                        pcm = bytes(ps["buffers"].pop(name))
+                        del ps["last_ts"][name]
+                        to_flush.append((name, pcm))
+
+            for name, pcm in to_flush:
+                if len(pcm) < _MIN_BYTES:
+                    continue
+                try:
+                    text = await loop.run_in_executor(
+                        None, _transcribe, ps, pcm,
+                    )
+                except Exception:
+                    log.exception("voice: transcription failed for %s", name)
+                    continue
+                if not text or text.strip("., ") == "":
+                    continue
+                log.info("voice: %s said: %s", name, text)
+                await bot.action("0", f"heard {name} say: {text}")
+    except asyncio.CancelledError:
+        pass
+    except Exception:
+        log.exception("voice: flush monitor error")
+
+
+# -- TTS: Piper fetch + playback --------------------------------------------
+
+
+def _fetch_tts(url: str) -> str | None:
+    """Fetch TTS WAV from Piper to a temp file.  Blocking."""
+    import tempfile
+    try:
+        resp = _urlopen(url, timeout=30, proxy=False)
+        data = resp.read()
+        resp.close()
+        if not data:
+            return None
+        tmp = tempfile.NamedTemporaryFile(
+            suffix=".wav", prefix="derp_tts_", delete=False,
+        )
+        tmp.write(data)
+        tmp.close()
+        return tmp.name
+    except Exception:
+        log.exception("voice: TTS fetch failed")
+        return None
+
+
+async def _tts_play(bot, text: str):
+    """Fetch TTS audio and play it via stream_audio."""
+    from pathlib import Path
+
+    ps = _ps(bot)
+    url = ps["piper_url"] + "?text=" + quote_plus(text)
+    loop = asyncio.get_running_loop()
+    wav_path = await loop.run_in_executor(None, _fetch_tts, url)
+    if wav_path is None:
+        return
+    try:
+        done = asyncio.Event()
+        await bot.stream_audio(str(wav_path), volume=1.0, on_done=done)
+        await done.wait()
+    finally:
+        Path(wav_path).unlink(missing_ok=True)
+
+
+# -- Listener lifecycle -----------------------------------------------------
+
+
+def _ensure_listener(bot):
+    """Register the sound listener callback (idempotent)."""
+    ps = _ps(bot)
+    if ps["_listener_registered"]:
+        return
+    if not hasattr(bot, "_sound_listeners"):
+        return
+    bot._sound_listeners.append(lambda user, chunk: _on_voice(bot, user, chunk))
+    ps["_listener_registered"] = True
+    log.info("voice: registered sound listener")
+
+
+def _ensure_flush_task(bot):
+    """Start the flush monitor if not running."""
+    ps = _ps(bot)
+    task = ps.get("flush_task")
+    if task and not task.done():
+        return
+    ps["flush_task"] = bot._spawn(
+        _flush_monitor(bot), name="voice-flush-monitor",
+    )
+
+
+def _stop_flush_task(bot):
+    """Cancel the flush monitor."""
+    ps = _ps(bot)
+    task = ps.get("flush_task")
+    if task and not task.done():
+        task.cancel()
+    ps["flush_task"] = None
+
+
+# -- Commands ----------------------------------------------------------------
+
+
+@command("listen", help="Voice: !listen [on|off] -- toggle STT", tier="admin")
+async def cmd_listen(bot, message):
+    """Toggle voice-to-text transcription."""
+    if not _is_mumble(bot):
+        await bot.reply(message, "Voice is Mumble-only")
+        return
+
+    ps = _ps(bot)
+    parts = message.text.split()
+    if len(parts) < 2:
+        state = "on" if ps["listen"] else "off"
+        await bot.reply(message, f"Listen: {state}")
+        return
+
+    sub = parts[1].lower()
+    if sub == "on":
+        ps["listen"] = True
+        _ensure_listener(bot)
+        _ensure_flush_task(bot)
+        await bot.reply(message, "Listening for voice")
+    elif sub == "off":
+        ps["listen"] = False
+        with ps["lock"]:
+            ps["buffers"].clear()
+            ps["last_ts"].clear()
+        _stop_flush_task(bot)
+        await bot.reply(message, "Stopped listening")
+    else:
+        await bot.reply(message, "Usage: !listen [on|off]")
+
+
+@command("say", help="Voice: !say <text> -- text-to-speech")
+async def cmd_say(bot, message):
+    """Speak text aloud via Piper TTS."""
+    if not _is_mumble(bot):
+        await bot.reply(message, "Voice is Mumble-only")
+        return
+
+    parts = message.text.split(None, 1)
+    if len(parts) < 2:
+        await bot.reply(message, "Usage: !say <text>")
+        return
+
+    text = parts[1].strip()
+    if len(text) > _MAX_SAY_LEN:
+        await bot.reply(message, f"Text too long (max {_MAX_SAY_LEN} chars)")
+        return
+
+    bot._spawn(_tts_play(bot, text), name="voice-tts")
+
+
+# -- Plugin lifecycle --------------------------------------------------------
+
+
+async def on_connected(bot) -> None:
+    """Re-register listener after reconnect if listen was on."""
+    if not _is_mumble(bot):
+        return
+    ps = _ps(bot)
+    if ps["listen"]:
+        _ensure_listener(bot)
+        _ensure_flush_task(bot)
diff --git a/tests/test_voice.py b/tests/test_voice.py
new file mode 100644
index 0000000..dc8d679
--- /dev/null
+++ b/tests/test_voice.py
@@ -0,0 +1,534 @@
+"""Tests for the voice STT/TTS plugin."""
+
+import asyncio
+import importlib.util
+import io
+import sys
+import time
+import wave
+from unittest.mock import AsyncMock, MagicMock, patch
+
+# -- Load plugin module directly ---------------------------------------------
+
+_spec = importlib.util.spec_from_file_location("voice", "plugins/voice.py")
+_mod = importlib.util.module_from_spec(_spec)
+sys.modules["voice"] = _mod
+_spec.loader.exec_module(_mod)
+
+
+# -- Fakes -------------------------------------------------------------------
+
+
+class _FakeState:
+    def __init__(self):
+        self._store: dict[str, dict[str, str]] = {}
+
+    def get(self, ns: str, key: str) -> str | None:
+        return self._store.get(ns, {}).get(key)
+
+    def set(self, ns: str, key: str, value: str) -> None:
+        self._store.setdefault(ns, {})[key] = value
+
+    def delete(self, ns: str, key: str) -> None:
+        self._store.get(ns, {}).pop(key, None)
+
+    def keys(self, ns: str) -> list[str]:
+        return list(self._store.get(ns, {}).keys())
+
+
+class _FakeBot:
+    """Minimal bot for voice plugin testing."""
+
+    def __init__(self, *, mumble: bool = True):
+        self.sent: list[tuple[str, str]] = []
+        self.replied: list[str] = []
+        self.actions: list[tuple[str, str]] = []
+        self.state = _FakeState()
+        self.config: dict = {}
+        self._pstate: dict = {}
+        self._tasks: set[asyncio.Task] = set()
+        self.nick = "derp"
+        self._sound_listeners: list = []
+        if mumble:
+            self.stream_audio = AsyncMock()
+
+    async def send(self, target: str, text: str) -> None:
+        self.sent.append((target, text))
+
+    async def reply(self, message, text: str) -> None:
+        self.replied.append(text)
+
+    async def action(self, target: str, text: str) -> None:
+        self.actions.append((target, text))
+
+    def _spawn(self, coro, *, name=None):
+        task = asyncio.ensure_future(coro)
+        self._tasks.add(task)
+        task.add_done_callback(self._tasks.discard)
+        return task
+
+
+class _Msg:
+    """Minimal message object."""
+
+    def __init__(self, text="!listen", nick="Alice", target="0",
+                 is_channel=True):
+        self.text = text
+        self.nick = nick
+        self.target = target
+        self.is_channel = is_channel
+        self.prefix = nick
+        self.command = "PRIVMSG"
+        self.params = [target, text]
+        self.tags = {}
+        self.raw = {}
+
+
+class _FakeSoundChunk:
+    """Minimal sound chunk with PCM data."""
+
+    def __init__(self, pcm: bytes = b"\x00\x00" * 960):
+        self.pcm = pcm
+
+
+# ---------------------------------------------------------------------------
+# TestMumbleGuard
+# ---------------------------------------------------------------------------
+
+
+class TestMumbleGuard:
+    def test_is_mumble_true(self):
+        bot = _FakeBot(mumble=True)
+        assert _mod._is_mumble(bot) is True
+
+    def test_is_mumble_false(self):
+        bot = _FakeBot(mumble=False)
+        assert _mod._is_mumble(bot) is False
+
+    def test_listen_non_mumble(self):
+        bot = _FakeBot(mumble=False)
+        msg = _Msg(text="!listen on")
+        asyncio.run(_mod.cmd_listen(bot, msg))
+        assert any("Mumble-only" in r for r in bot.replied)
+
+    def test_say_non_mumble(self):
+        bot = _FakeBot(mumble=False)
+        msg = _Msg(text="!say hello")
+        asyncio.run(_mod.cmd_say(bot, msg))
+        assert any("Mumble-only" in r for r in bot.replied)
+
+
+# ---------------------------------------------------------------------------
+# TestListenCommand
+# ---------------------------------------------------------------------------
+
+
+class TestListenCommand:
+    def test_listen_status(self):
+        bot = _FakeBot()
+        msg = _Msg(text="!listen")
+        asyncio.run(_mod.cmd_listen(bot, msg))
+        assert any("off" in r.lower() for r in bot.replied)
+
+    def test_listen_on(self):
+        bot = _FakeBot()
+        msg = _Msg(text="!listen on")
+        asyncio.run(_mod.cmd_listen(bot, msg))
+        ps = _mod._ps(bot)
+        assert ps["listen"] is True
+        assert any("Listening" in r for r in bot.replied)
+
+    def test_listen_off(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        ps["buffers"]["Alice"] = bytearray(b"\x00" * 100)
+        ps["last_ts"]["Alice"] = time.monotonic()
+        msg = _Msg(text="!listen off")
+        asyncio.run(_mod.cmd_listen(bot, msg))
+        assert ps["listen"] is False
+        assert ps["buffers"] == {}
+        assert ps["last_ts"] == {}
+        assert any("Stopped" in r for r in bot.replied)
+
+    def test_listen_invalid(self):
+        bot = _FakeBot()
+        msg = _Msg(text="!listen maybe")
+        asyncio.run(_mod.cmd_listen(bot, msg))
+        assert any("Usage" in r for r in bot.replied)
+
+
+# ---------------------------------------------------------------------------
+# TestSayCommand
+# ---------------------------------------------------------------------------
+
+
+class TestSayCommand:
+    def test_say_no_text(self):
+        bot = _FakeBot()
+        msg = _Msg(text="!say")
+        asyncio.run(_mod.cmd_say(bot, msg))
+        assert any("Usage" in r for r in bot.replied)
+
+    def test_say_too_long(self):
+        bot = _FakeBot()
+        text = "x" * 501
+        msg = _Msg(text=f"!say {text}")
+        asyncio.run(_mod.cmd_say(bot, msg))
+        assert any("too long" in r.lower() for r in bot.replied)
+
+    def test_say_spawns_task(self):
+        bot = _FakeBot()
+        msg = _Msg(text="!say hello world")
+
+        spawned = []
+        original_spawn = bot._spawn
+
+        def track_spawn(coro, *, name=None):
+            spawned.append(name)
+            coro.close()
+            task = MagicMock()
+            task.done.return_value = False
+            return task
+
+        bot._spawn = track_spawn
+        asyncio.run(_mod.cmd_say(bot, msg))
+        assert "voice-tts" in spawned
+
+
+# ---------------------------------------------------------------------------
+# TestAudioBuffering
+# ---------------------------------------------------------------------------
+
+
+class TestAudioBuffering:
+    def test_accumulates_pcm(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        user = {"name": "Alice"}
+        chunk = _FakeSoundChunk(b"\x01\x02" * 480)
+        _mod._on_voice(bot, user, chunk)
+        assert "Alice" in ps["buffers"]
+        assert len(ps["buffers"]["Alice"]) == 960
+
+    def test_ignores_own_nick(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        user = {"name": "derp"}
+        chunk = _FakeSoundChunk(b"\x01\x02" * 480)
+        _mod._on_voice(bot, user, chunk)
+        assert "derp" not in ps["buffers"]
+
+    def test_respects_listen_false(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = False
+        user = {"name": "Alice"}
+        chunk = _FakeSoundChunk(b"\x01\x02" * 480)
+        _mod._on_voice(bot, user, chunk)
+        assert ps["buffers"] == {}
+
+    def test_caps_at_max_bytes(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        user = {"name": "Alice"}
+        # Fill beyond max
+        big_chunk = _FakeSoundChunk(b"\x00\x01" * (_mod._MAX_BYTES // 2 + 100))
+        _mod._on_voice(bot, user, big_chunk)
+        assert len(ps["buffers"]["Alice"]) <= _mod._MAX_BYTES
+
+    def test_empty_pcm_ignored(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        user = {"name": "Alice"}
+        chunk = _FakeSoundChunk(b"")
+        _mod._on_voice(bot, user, chunk)
+        assert "Alice" not in ps["buffers"]
+
+    def test_none_user_ignored(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        chunk = _FakeSoundChunk(b"\x01\x02" * 480)
+        _mod._on_voice(bot, "not_a_dict", chunk)
+        assert ps["buffers"] == {}
+
+    def test_updates_timestamp(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        user = {"name": "Alice"}
+        chunk = _FakeSoundChunk(b"\x01\x02" * 480)
+        _mod._on_voice(bot, user, chunk)
+        assert "Alice" in ps["last_ts"]
+        ts1 = ps["last_ts"]["Alice"]
+        _mod._on_voice(bot, user, chunk)
+        assert ps["last_ts"]["Alice"] >= ts1
+
+
+# ---------------------------------------------------------------------------
+# TestFlushLogic
+# ---------------------------------------------------------------------------
+
+
+class TestFlushLogic:
+    def test_silence_gap_triggers_flush(self):
+        """Buffer is flushed and transcribed after silence gap."""
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        ps["silence_gap"] = 0.1  # very short for testing
+
+        # Pre-populate buffer with enough PCM (> _MIN_BYTES)
+        pcm = b"\x00\x01" * (_mod._MIN_BYTES // 2 + 100)
+        with ps["lock"]:
+            ps["buffers"]["Alice"] = bytearray(pcm)
+            ps["last_ts"]["Alice"] = time.monotonic() - 1.0  # already silent
+
+        async def _check():
+            with patch.object(_mod, "_transcribe", return_value="hello"):
+                task = asyncio.create_task(_mod._flush_monitor(bot))
+                await asyncio.sleep(1.0)
+                ps["listen"] = False  # stop the monitor
+                await asyncio.sleep(0.2)
+                try:
+                    await asyncio.wait_for(task, timeout=2)
+                except (asyncio.CancelledError, asyncio.TimeoutError):
+                    pass
+            assert any("hello" in a[1] for a in bot.actions)
+
+        asyncio.run(_check())
+
+    def test_min_duration_filter(self):
+        """Short utterances (< _MIN_BYTES) are discarded."""
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        ps["silence_gap"] = 0.1
+
+        # Buffer too small
+        with ps["lock"]:
+            ps["buffers"]["Alice"] = bytearray(b"\x00\x01" * 10)
+            ps["last_ts"]["Alice"] = time.monotonic() - 1.0
+
+        async def _check():
+            with patch.object(_mod, "_transcribe", return_value="x") as mock_t:
+                task = asyncio.create_task(_mod._flush_monitor(bot))
+                await asyncio.sleep(0.5)
+                ps["listen"] = False
+                await asyncio.sleep(0.2)
+                try:
+                    await asyncio.wait_for(task, timeout=2)
+                except (asyncio.CancelledError, asyncio.TimeoutError):
+                    pass
+            mock_t.assert_not_called()
+
+        asyncio.run(_check())
+
+    def test_buffer_cleared_after_flush(self):
+        """Buffer and timestamp are removed after flushing."""
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        ps["silence_gap"] = 0.1
+
+        pcm = b"\x00\x01" * (_mod._MIN_BYTES // 2 + 100)
+        with ps["lock"]:
+            ps["buffers"]["Alice"] = bytearray(pcm)
+            ps["last_ts"]["Alice"] = time.monotonic() - 1.0
+
+        async def _check():
+            with patch.object(_mod, "_transcribe", return_value="test"):
+                task = asyncio.create_task(_mod._flush_monitor(bot))
+                await asyncio.sleep(0.5)
+                ps["listen"] = False
+                await asyncio.sleep(0.2)
+                try:
+                    await asyncio.wait_for(task, timeout=2)
+                except (asyncio.CancelledError, asyncio.TimeoutError):
+                    pass
+            assert "Alice" not in ps["buffers"]
+            assert "Alice" not in ps["last_ts"]
+
+        asyncio.run(_check())
+
+
+# ---------------------------------------------------------------------------
+# TestPcmToWav
+# ---------------------------------------------------------------------------
+
+
+class TestPcmToWav:
+    def test_valid_wav(self):
+        pcm = b"\x00\x00" * 48000  # 1 second of silence
+        wav_data = _mod._pcm_to_wav(pcm)
+        # Should start with RIFF header
+        assert wav_data[:4] == b"RIFF"
+        # Parse it back
+        buf = io.BytesIO(wav_data)
+        with wave.open(buf, "rb") as wf:
+            assert wf.getnchannels() == 1
+            assert wf.getsampwidth() == 2
+            assert wf.getframerate() == 48000
+            assert wf.getnframes() == 48000
+
+    def test_empty_pcm(self):
+        wav_data = _mod._pcm_to_wav(b"")
+        buf = io.BytesIO(wav_data)
+        with wave.open(buf, "rb") as wf:
+            assert wf.getnframes() == 0
+
+
+# ---------------------------------------------------------------------------
+# TestTranscribe
+# ---------------------------------------------------------------------------
+
+
+class TestTranscribe:
+    def test_parse_json_response(self):
+        ps = {"whisper_url": "http://localhost:8080/inference"}
+        pcm = b"\x00\x00" * 4800  # 0.1s
+        resp = MagicMock()
+        resp.read.return_value = b'{"text": "hello world"}'
+        with patch.object(_mod, "_urlopen", return_value=resp):
+            text = _mod._transcribe(ps, pcm)
+        assert text == "hello world"
+
+    def test_empty_text(self):
+        ps = {"whisper_url": "http://localhost:8080/inference"}
+        pcm = b"\x00\x00" * 4800
+        resp = MagicMock()
+        resp.read.return_value = b'{"text": ""}'
+        with patch.object(_mod, "_urlopen", return_value=resp):
+            text = _mod._transcribe(ps, pcm)
+        assert text == ""
+
+    def test_missing_text_key(self):
+        ps = {"whisper_url": "http://localhost:8080/inference"}
+        pcm = b"\x00\x00" * 4800
+        resp = MagicMock()
+        resp.read.return_value = b'{"result": "something"}'
+        with patch.object(_mod, "_urlopen", return_value=resp):
+            text = _mod._transcribe(ps, pcm)
+        assert text == ""
+
+
+# ---------------------------------------------------------------------------
+# TestPerBotState
+# ---------------------------------------------------------------------------
+
+
+class TestPerBotState:
+    def test_ps_initializes(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        assert ps["listen"] is False
+        assert ps["buffers"] == {}
+        assert ps["last_ts"] == {}
+
+    def test_ps_stable_reference(self):
+        bot = _FakeBot()
+        ps1 = _mod._ps(bot)
+        ps2 = _mod._ps(bot)
+        assert ps1 is ps2
+
+    def test_ps_isolated_per_bot(self):
+        bot1 = _FakeBot()
+        bot2 = _FakeBot()
+        _mod._ps(bot1)["listen"] = True
+        assert _mod._ps(bot2)["listen"] is False
+
+    def test_ps_config_override(self):
+        bot = _FakeBot()
+        bot.config = {"voice": {"silence_gap": 3.0}}
+        ps = _mod._ps(bot)
+        assert ps["silence_gap"] == 3.0
+
+
+# ---------------------------------------------------------------------------
+# TestEnsureListener
+# ---------------------------------------------------------------------------
+
+
+class TestEnsureListener:
+    def test_registers_callback(self):
+        bot = _FakeBot()
+        _mod._ps(bot)  # init state
+        _mod._ensure_listener(bot)
+        assert len(bot._sound_listeners) == 1
+        ps = _mod._ps(bot)
+        assert ps["_listener_registered"] is True
+
+    def test_idempotent(self):
+        bot = _FakeBot()
+        _mod._ps(bot)
+        _mod._ensure_listener(bot)
+        _mod._ensure_listener(bot)
+        assert len(bot._sound_listeners) == 1
+
+    def test_no_listener_without_attr(self):
+        bot = _FakeBot()
+        del bot._sound_listeners
+        _mod._ps(bot)
+        _mod._ensure_listener(bot)
+        # Should not raise, just skip
+
+    def test_callback_calls_on_voice(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+        _mod._ensure_listener(bot)
+        user = {"name": "Alice"}
+        chunk = _FakeSoundChunk(b"\x01\x02" * 480)
+        bot._sound_listeners[0](user, chunk)
+        assert "Alice" in ps["buffers"]
+
+
+# ---------------------------------------------------------------------------
+# TestOnConnected
+# ---------------------------------------------------------------------------
+
+
+class TestOnConnected:
+    def test_reregisters_when_listening(self):
+        bot = _FakeBot()
+        ps = _mod._ps(bot)
+        ps["listen"] = True
+
+        spawned = []
+
+        def fake_spawn(coro, *, name=None):
+            task = MagicMock()
+            task.done.return_value = False
+            spawned.append(name)
+            coro.close()
+            return task
+
+        bot._spawn = fake_spawn
+        asyncio.run(_mod.on_connected(bot))
+        assert ps["_listener_registered"] is True
+        assert "voice-flush-monitor" in spawned
+
+    def test_noop_when_not_listening(self):
+        bot = _FakeBot()
+        _mod._ps(bot)  # init but listen=False
+
+        spawned = []
+
+        def fake_spawn(coro, *, name=None):
+            spawned.append(name)
+            coro.close()
+            return MagicMock()
+
+        bot._spawn = fake_spawn
+        asyncio.run(_mod.on_connected(bot))
+        assert "voice-flush-monitor" not in spawned
+
+    def test_noop_non_mumble(self):
+        bot = _FakeBot(mumble=False)
+        asyncio.run(_mod.on_connected(bot))
+        # Should not raise or register anything