Beddington

A privacy-first baby cry monitor that runs entirely on your own device. No cloud, no streaming, no account.

  

   

---

Most baby monitors send audio and video to a vendor's cloud. Beddington keeps everything on the device in front of you. It listens for sustained crying, logs each episode, writes a plain-language night log and a morning digest, and can play one soothing sound before it pings you. Raw audio and video never leave the device. It is an assistive notebook for tired parents, not a medical device, and it runs with every cloud feature switched off.

Quickstart · How it works · Features · Configuration · Privacy

Why Beddington exists

You want to know when your baby is crying. You do not want a microphone in the nursery uploading your child's sounds to someone else's servers, on someone else's terms, with someone else's retention policy.

Beddington runs the cry detection on your own hardware, a Raspberry Pi or a laptop. The audio is analysed on the device and thrown away. What you keep is a short event record, a readable night log, and a morning digest. Nothing about your night leaves the room unless you explicitly turn on the optional text-only digest polish.

What it does today

What	How
On-device cry detection	The official YAMNet TFLite "Baby cry, infant cry" model, run locally on your audio
Fewer false alarms	A confidence threshold, sustained-duration debounce, release delay, and notification cooldown you can tune
Readable night log + morning digest	Plain-text night-log.txt and morning-digest.txt, plus a structured events.json
One soothing sound, your choice	Optional Tier 1 plays one selected sound or music preset before it notifies you
Quiet checks while soothing	Pauses, listens, and requires repeated quiet readings before it logs that crying has stopped
Laptop or Raspberry Pi	A .wav file or a live USB microphone runs the same detector and state machine
Private by construction	Raw audio and video never leave the device; cloud features are off by default
Optional LLM digest polish	Only derived event text is sent, and only when you pass --llm

Quickstart

Use Python 3.11 to 3.14.

git clone https://github.com/moelzek/Beddington.git
cd Beddington
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install ".[dev]"

Point analyze at any short .wav of a crying baby. A 10 to 30 second clip is plenty. The development sample recording is kept out of this public repo, so bring your own audio file:

beddington --config config/default.toml analyze path/to/crying.wav \
  --output output/sample-night

The first run downloads and verifies the official 3.9 MB YAMNet TFLite model into ~/.cache/beddington/models/, then processes the audio locally. You should see something like:

[Beddington] Sustained crying detected (...). Please check Rayan.
Beddington detected 1 sustained crying episode...
Events: output/sample-night/events.json
Readable log: output/sample-night/night-log.txt
Morning digest: output/sample-night/morning-digest.txt

Open the three output files:

cat output/sample-night/night-log.txt
cat output/sample-night/morning-digest.txt
python -m json.tool output/sample-night/events.json

Generated output is gitignored.

How it works

  microphone or .wav            YAMNet (on-device)          deterministic state machine            you
 ───────────────────►  cry score  ───────────────►  threshold · debounce · cooldown  ───────►  night log
                                                            + optional soothe preset             morning digest
                                                                                                 notification

1. Audio comes from a .wav file or a live 16 kHz mono microphone window.
2. The local YAMNet model returns a baby-cry score for each window. The audio is then discarded.
3. A deterministic state machine decides what counts as a real episode: the score must clear a threshold, stay high for a sustained duration, and respect a release delay and a cooldown between notifications.
4. Optionally, Beddington plays one configured soothe preset and runs quiet checks before it escalates to a notification.
5. You get an events.json, a readable night-log.txt, and a morning-digest.txt.

Detection and timing are deterministic. The optional language model only ever rewrites the final text digest, and only if you ask it to.

More commands

# Listen on a live microphone for 60 seconds (needs the mic extra)
python -m pip install ".[mic]"
beddington --config config/default.toml listen --seconds 60 --output output/live

# Try the Tier 1 soothe preset in dry-run mode (records the preset, no sound)
beddington --config config/tier1-demo.toml analyze path/to/crying.wav --output output/tier1-demo

# Preview the selected soothe sound through your speaker, briefly
beddington --config config/default.toml preview-soothe --seconds 5

# Run the hardware-free test suite (no model download, no mic, no API key)
python -m pytest

On Raspberry Pi OS the microphone path may also need sudo apt install libportaudio2.

Beddington also has bench-only, local-only camera utilities (camera-smoke, visual-change, camera-change) that write derived metrics and delete raw frames by default. They do not run nursery video, and video never triggers or suppresses a notification.

Configuration

Everything lives in config/default.toml. The knobs that matter most for false alarms:

Setting	What it does	Default
threshold	Minimum YAMNet baby-cry score to count	0.40
sustained_seconds	How long the score must stay high before an event	1.5
release_seconds	How long it must stay low before the episode ends	—
notification_cooldown_seconds	Minimum time between notifications	—

YAMNet scores are uncalibrated model scores, not probabilities. Tune them against recordings from your own room before you rely on notifications. Soothe behaviour (soothe.enabled, soothe.preset, soothe.player, and the quiet_check block) is documented in the same file.

The bundled soothe sounds in assets/soothe/ are local audition assets for testing the preset path and dashboard controls. The womb-like file is research/audition material, not medical or safety evidence, and not evidence that any sound will settle a given baby.

What's inside

src/beddington/    application code
tests/             hardware-free tests
config/            deterministic thresholds and feature flags
assets/soothe/     local soothe sounds and dashboard catalog
output/            generated logs (gitignored)

Privacy and safety

• Raw audio and video never leave the device. Only derived event text is ever sent anywhere, and only when you pass --llm.
• Beddington does not diagnose illness and does not detect SIDS, apnoea, or fever. It is not a medical device and does not replace adult supervision or approved monitoring equipment.
• Uncertain interpretations are labelled best guess.
• Keep the companion beside the cot, never in it, and keep hot compute in a vented base.
• The complete app works with every cloud feature disabled. Never commit your .env.

---

Built by Mo Elzek

If Beddington is useful to you, star it so other parents can find it.