Cap buffered audio and keep transcripts out of INFO logs

Wyoming has no authentication, so buffered audio is attacker-controlled.
self.audio grew until AudioStop with no bound: measured at ~11 MB/s over
loopback, one connection exhausts 32 GB in under an hour, and a stuck
satellite that never sends AudioStop does the same by accident. Cap it at
--max-audio-seconds (default 120), dropping the excess with a single warning
while still transcribing what was captured. Verified: a client streaming
10.8 GB now moves server RSS by 213 MB rather than 10.8 GB.

Transcripts were logged at INFO. Log files are long-lived and world-readable
under /tmp on macOS, so every voice command sat in plaintext readable by any
local account. INFO now records duration, latency and character count; the
text moved behind --debug.

Both are covered by mutation-checked tests, and the README gains a Security
section covering the unauthenticated trust boundary, the 0.0.0.0 bind that
also exposes VPN interfaces, and running the daemon as a non-admin user.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
2026-07-29 03:44:25 +01:00
co-authored by Claude Opus 5
parent 9bf887e780
commit 5a6fc62106
4 changed files with 170 additions and 9 deletions
+41
View File
@@ -302,6 +302,47 @@ tail -f /tmp/local.wyoming-parakeet.stderr
Each request logs audio duration, inference time and the transcript.
## Security
**The Wyoming protocol has no authentication or transport encryption.** Any
client that can reach the port can submit audio, and Home Assistant trusts
whatever this service returns — a transcript becomes an intent, and an intent
unlocks doors. Treat the port as a trust boundary and keep it on a network you
control.
Two things this server does about that:
- **Buffered audio is capped** (`--max-audio-seconds`, default 120). Audio
accumulates until `AudioStop`, and a client that never sends one — malicious
or just stuck — otherwise grows the buffer indefinitely. Measured at ~11 MB/s
over loopback, enough to exhaust 32 GB in under an hour from one connection.
Over the cap, further audio is dropped with a single warning and whatever was
captured is still transcribed.
- **Transcripts are not logged at INFO.** The log records duration, latency and
character count; the text itself is behind `--debug`. Log files are
long-lived, and on macOS `/tmp` they are world-readable by default — meaning
every voice command would otherwise sit in plaintext readable by any local
account.
Worth doing yourself, depending on your threat model:
- **Bind to one interface.** The daemon listens on `0.0.0.0`, so it is exposed
on every network the host is attached to — including VPN interfaces like
Tailscale, which is easy to overlook. Set the URI to a specific address
(`--uri tcp://192.168.1.10:7892`) or firewall the port to your Home
Assistant host.
- **Do not run it as an admin account.** The installer defaults `--user` to
whoever runs it. On a typical macOS setup that account is in `admin`, and if
`%admin` has a `NOPASSWD` sudo rule then a compromise of this service is a
direct path to root. A dedicated non-admin service account costs nothing.
- **Pin your dependencies** if you care about supply chain. `requirements.txt`
is deliberately loose so `install.sh` picks up fixes; pin exact versions (and
ideally hashes) if you would rather audit upgrades. Model weights are
`safetensors`, so loading them does not execute code, but the initial
download from HuggingFace is trust-on-first-use — pin a revision if that
matters to you. `HF_HUB_OFFLINE=1` in the daemon means it never re-fetches
after that point.
## License
MIT
+88 -4
View File
@@ -1,4 +1,5 @@
"""Tests for the Wyoming event handling."""
from argparse import Namespace
from unittest.mock import AsyncMock, MagicMock
import pytest
@@ -7,7 +8,12 @@ from wyoming.audio import AudioChunk, AudioStop
from wyoming.info import AsrModel, AsrProgram, Attribution, Describe, Info
from wyoming_parakeet.engine import SAMPLE_RATE
from wyoming_parakeet.handler import ParakeetEventHandler
from wyoming_parakeet.handler import (
DEFAULT_MAX_AUDIO_SECONDS,
ParakeetEventHandler,
)
ARGS = Namespace(max_audio_seconds=DEFAULT_MAX_AUDIO_SECONDS)
INFO = Info(
asr=[
@@ -36,7 +42,7 @@ INFO = Info(
def handler():
engine = MagicMock()
engine.transcribe = AsyncMock(return_value="turn off the kitchen lights")
h = ParakeetEventHandler(INFO, MagicMock(), engine, MagicMock(), MagicMock())
h = ParakeetEventHandler(INFO, ARGS, engine, MagicMock(), MagicMock())
h.write_event = AsyncMock()
return h
@@ -172,8 +178,8 @@ async def test_concurrent_handlers_do_not_share_audio():
seen = []
engine.transcribe.side_effect = lambda pcm: seen.append(pcm) or ""
a = ParakeetEventHandler(INFO, MagicMock(), engine, MagicMock(), MagicMock())
b = ParakeetEventHandler(INFO, MagicMock(), engine, MagicMock(), MagicMock())
a = ParakeetEventHandler(INFO, ARGS, engine, MagicMock(), MagicMock())
b = ParakeetEventHandler(INFO, ARGS, engine, MagicMock(), MagicMock())
a.write_event = AsyncMock()
b.write_event = AsyncMock()
@@ -224,3 +230,81 @@ async def test_8bit_input_is_widened(handler):
await handler.handle_event(AudioStop().event())
assert len(handler.engine.transcribe.await_args[0][0]) == SAMPLE_RATE * 2
# --- resource limits --------------------------------------------------------
# Wyoming is unauthenticated, so buffered audio is attacker-controlled.
def capped_handler(seconds):
engine = MagicMock()
engine.transcribe = AsyncMock(return_value="ok")
h = ParakeetEventHandler(
INFO, Namespace(max_audio_seconds=seconds), engine, MagicMock(), MagicMock()
)
h.write_event = AsyncMock()
return h
async def test_audio_buffer_is_capped():
handler = capped_handler(1.0)
for _ in range(10):
await handler.handle_event(chunk(SAMPLE_RATE)) # 10s into a 1s cap
assert len(handler.audio) == SAMPLE_RATE * 2
async def test_truncated_audio_is_still_transcribed():
"""Cap the memory, don't drop the user's command on the floor."""
handler = capped_handler(1.0)
for _ in range(5):
await handler.handle_event(chunk(SAMPLE_RATE))
await handler.handle_event(AudioStop().event())
handler.engine.transcribe.assert_awaited_once()
assert len(handler.engine.transcribe.await_args[0][0]) == SAMPLE_RATE * 2
assert Transcript.from_event(handler.write_event.await_args[0][0]).text == "ok"
async def test_truncation_warns_once(caplog):
handler = capped_handler(1.0)
with caplog.at_level("WARNING"):
for _ in range(6):
await handler.handle_event(chunk(SAMPLE_RATE))
assert caplog.text.count("Audio exceeded") == 1
async def test_cap_resets_between_utterances():
handler = capped_handler(1.0)
for _ in range(3):
await handler.handle_event(chunk(SAMPLE_RATE))
await handler.handle_event(AudioStop().event())
assert handler.truncated is False
await handler.handle_event(chunk(SAMPLE_RATE // 2))
assert len(handler.audio) == SAMPLE_RATE # accepted again, not still capped
async def test_audio_under_the_cap_is_untouched():
handler = capped_handler(DEFAULT_MAX_AUDIO_SECONDS)
await handler.handle_event(chunk(SAMPLE_RATE * 3))
assert len(handler.audio) == SAMPLE_RATE * 3 * 2
# --- transcript privacy -----------------------------------------------------
# Logs are long-lived and readable by other local accounts.
async def test_transcript_text_is_not_logged_at_info(handler, caplog):
handler.engine.transcribe = AsyncMock(return_value="unlock the front door")
with caplog.at_level("INFO"):
await handler.handle_event(chunk(SAMPLE_RATE))
await handler.handle_event(AudioStop().event())
assert "unlock the front door" not in caplog.text
async def test_transcript_text_is_available_at_debug(handler, caplog):
handler.engine.transcribe = AsyncMock(return_value="unlock the front door")
with caplog.at_level("DEBUG"):
await handler.handle_event(chunk(SAMPLE_RATE))
await handler.handle_event(AudioStop().event())
assert "unlock the front door" in caplog.text
+12 -2
View File
@@ -8,7 +8,7 @@ from wyoming.info import AsrModel, AsrProgram, Attribution, Info
from wyoming.server import AsyncServer
from .engine import ParakeetEngine
from .handler import ParakeetEventHandler
from .handler import DEFAULT_MAX_AUDIO_SECONDS, ParakeetEventHandler
_LOGGER = logging.getLogger(__name__)
__version__ = "1.0.0"
@@ -25,7 +25,17 @@ def build_parser() -> argparse.ArgumentParser:
parser.add_argument("--uri", required=True, help="unix:// or tcp://")
parser.add_argument("--model", default=DEFAULT_MODEL, help="HuggingFace model id")
parser.add_argument("--language", default="en", help="Language code reported to HA")
parser.add_argument("--debug", action="store_true", help="Log DEBUG messages")
parser.add_argument(
"--max-audio-seconds",
type=float,
default=DEFAULT_MAX_AUDIO_SECONDS,
help="Cap on buffered audio per utterance (default: %(default)s)",
)
parser.add_argument(
"--debug",
action="store_true",
help="Log DEBUG messages, including transcript text",
)
parser.add_argument("--log-format", default=logging.BASIC_FORMAT)
parser.add_argument("--version", action="version", version=__version__)
return parser
+29 -3
View File
@@ -12,6 +12,13 @@ from .engine import SAMPLE_RATE
_LOGGER = logging.getLogger(__name__)
# Wyoming has no authentication, so any client that can reach the port can
# stream audio. Without a cap, self.audio grows until AudioStop -- measured at
# ~11 MB/s over loopback, which exhausts 32 GB in under an hour from a single
# connection. A stuck satellite that never sends AudioStop does the same thing
# by accident. No real voice command approaches this bound.
DEFAULT_MAX_AUDIO_SECONDS = 120
class ParakeetEventHandler(AsyncEventHandler):
def __init__(self, wyoming_info: Info, cli_args, engine, *args, **kwargs):
@@ -20,13 +27,27 @@ class ParakeetEventHandler(AsyncEventHandler):
self.wyoming_info_event = wyoming_info.event()
self.engine = engine
self.audio = bytes()
self.truncated = False
self.converter = AudioChunkConverter(rate=SAMPLE_RATE, width=2, channels=1)
max_seconds = getattr(cli_args, "max_audio_seconds", DEFAULT_MAX_AUDIO_SECONDS)
self.max_bytes = int(max_seconds * SAMPLE_RATE * 2)
def _append(self, chunk: bytes) -> None:
room = self.max_bytes - len(self.audio)
if room > 0:
self.audio += chunk[:room]
if len(self.audio) >= self.max_bytes and not self.truncated:
self.truncated = True
_LOGGER.warning(
"Audio exceeded %.0fs; ignoring the rest of this utterance",
self.max_bytes / (SAMPLE_RATE * 2),
)
async def handle_event(self, event: Event) -> bool:
if AudioChunk.is_type(event.type):
if not self.audio:
_LOGGER.debug("Receiving audio")
self.audio += self.converter.convert(AudioChunk.from_event(event)).audio
self._append(self.converter.convert(AudioChunk.from_event(event)).audio)
return True
if AudioStop.is_type(event.type):
@@ -34,12 +55,16 @@ class ParakeetEventHandler(AsyncEventHandler):
started = time.monotonic()
try:
text = await self.engine.transcribe(self.audio)
# The transcript is everything the user said, and the log file
# is long-lived and readable by other local accounts. Keep the
# operational signal at INFO and the content behind --debug.
_LOGGER.info(
"%.2fs audio -> %.0fms :: %r",
"%.2fs audio -> %.0fms, %d chars",
duration,
(time.monotonic() - started) * 1000,
text,
len(text),
)
_LOGGER.debug("Transcript: %r", text)
except Exception:
# wyoming's run loop has no except clause, so letting this
# propagate closes the connection without ever sending a
@@ -49,6 +74,7 @@ class ParakeetEventHandler(AsyncEventHandler):
text = ""
finally:
self.audio = bytes()
self.truncated = False
await self.write_event(Transcript(text=text).event())
return False