Sonnette - Smart SIP Doorbell powered by RK3506

A smart doorbell that runs Linux and acts as its own phone server. When someone presses the button, it calls your phone: a regular phone call, not a push notification from some app.

The doorbell is a SIP client (like a VoIP phone) built around pjsua/pjsip, a battle-tested open-source SIP stack. It registers on a Twilio SIP domain, and when triggered, it asks Twilio to call your number and bridge the audio. You pick up the phone, and talk to the person at the door directly, with no cloud/IoT service in-between.

Twilio is just the SIP-to-PSTN bridge. The audio goes through Twilio's infrastructure during the call, but nothing is stored or recorded. You can replace Twilio with any SIP provider (OVH Telecom, VoIP.ms, etc.) if you prefer.

• Running cost: ~1.20 €/month for the phone number (French 09xx) plus ~0.04 €/min per call to a local mobile. See docs/twilio.md for details.

• Build costs: ~50 € PCBA (JLCPCB, single qty) plus a dozen euros for the speaker, 3D-printed enclosure & WiFi dongle.

Architecture

Two processes on the board: sonnette (Go) orchestrates everything, pjsua (C/pjsip) handles SIP and audio.

Button → sonnette → Twilio API →  phone rings
            │                          ↓
            │                     you pick up
            │                          ↓
            └→ aplay chime       Twilio bridges
               on speaker         SIP ←→ PSTN
                                       ↓
                         Continuous bidirectional stream:
                      PDM mic → pjsua → RTP → Twilio → phone
                      speaker ← pjsua ← RTP ← Twilio ← phone

sonnette (Go daemon)

The main process. Source in sonnette/.

• GPIO watcher: monitors the ring button via sysfs poll() with software debounce (300 ms). Active low, falling edge.
• Dingdong playback: plays dingdong.wav on the speaker via aplay (synchronous, blocks until done to avoid ALSA device contention with pjsua).
• Twilio REST API: POSTs to Twilio to initiate the outbound call to your phone. 10 s HTTP timeout to avoid blocking on flaky internet.
• Cooldown: prevents button spam (configurable: default 30 s cooldown, max 10 calls/hour).
• pjsua supervisor: starts pjsua as a child process, restarts it on crash (3 s backoff). SIP credentials are written to a temp file (/run/pjsua-account.cfg, chmod 0600) so they don't appear in ps. pjsua's stdout/stderr is captured to /var/log/pjsua.log via stdbuf -oL for line-buffered output.
• Health endpoint: serves GET /health on a configurable port (default 8080) returning JSON with uptime, pjsua status, WiFi state, and last ring activity. Used by a home server to detect when the doorbell is down.
• Event webhook: optional. POSTs JSON events (button_press, call_initiated, call_failed, periodic heartbeat) to a configurable URL for push-style notifications.
• Config: reads sonnette.yaml (settings) and sonnette.conf (credentials, shell KEY="value" format, parsed directly in Go).

pjsua (SIP client)

Cross-compiled C binary from pjproject. Source/Makefile in pjsua/.

Handles SIP registration on Twilio, auto-answers incoming INVITEs, and bridges two-way audio between the PDM mic + I2S speaker and the RTP stream. Runs at 16 kHz hardware sample rate, resampled internally to 8 kHz PCMU for SIP.

Twilio

SIP-to-PSTN bridge. Connects the VoIP call to your phone number. See docs/twilio.md for the full setup guide.

Quick start

1. Set up Twilio — account, phone number, SIP domain, TwiML Bin. You need the credentials in hand before deploying. See docs/twilio.md.
2. Build the Linux image — copy the BSP files into the Rockchip SDK, build the SDK image, then flash it. This may happen directly on your host or inside the SDK's Docker container depending on your setup; see docs/hardware.md.
3. Deploy the software — build sonnette + pjsua, sanity-check the fresh image audio path, prepare secrets, deploy over ADB, install init supervision, harden SSH, then verify the runtime. See docs/deploy.md.
4. Press the button — phone rings.

The repo uses a small, readable Makefile to keep repetitive steps simple without hiding the system. The .md guides explain the workflow and trade-offs; make help lists the available shortcuts; and the Makefile itself is the place to inspect the exact commands.

Hardware

Developed on a custom Rockchip RK3506G2 board. The bsp/ directory has the device tree and SDK configs; the eda/ directory contains the EasyEDA Pro design files. The software itself is plain Linux + Go + pjsua, it runs on any SBC with ALSA audio and a GPIO (Raspberry Pi, BeagleBone, etc.).

The pictured enclosure is a 3D-printed prototype (its CAD sources are not published here yet).

To port it, expect to adjust:

• Init system: the deploy scripts assume BusyBox inittab. On a systemd distro, use a sonnette.service unit instead. Logs work the same with journalctl -u sonnette.
• ALSA card numbers: deploy/asound.conf references the RK3506 PDM (hw:0,0) and SAI1 (hw:1,0) cards. Other boards expose audio on different cards; adjust the device names and the mono→stereo routing if your codec doesn't need it.
• PDM gain: amixer -c 0 cset numid=7 100% is RK3506-specific. Other mics/codecs expose different mixer controls (or none).
• Network interface: defaults to wlan0; rename in S90wifi if your board uses something else.

See docs/hardware.md for the pin map, BSP build and hardware reference, and docs/deploy.md for the ALSA, pjsua and WiFi runbooks. Cross-compile of pjsua against the target sysroot is documented in pjsua/README.md.

Security and privacy model

This is a personal/home device, not an industrial high-security product. The production setup does the practical basics: disable ADB, harden SSH to key-only login, randomize the root password, optionally disable UART login, keep credentials in 0600 files, optionally enable a restrictive firewall, and expose health/webhook signals so failures or tampering can be noticed quickly.

The design still assumes physical access is a strong attacker capability. The software runs as root on a small Buildroot system, and if someone removes the doorbell from the wall, they could theoretically extract WiFi and Twilio/SIP credentials from flash. If USB ADB and UART login are disabled, normal maintenance goes through SSH; if SSH/WiFi is lost, recovery is Maskrom reflash. The expected security response to a missing or compromised unit is operational: revoke the SIP password/Auth Token and rotate WiFi credentials if needed.

That trade-off is intentional. The project avoids an always-on video doorbell and avoids a proprietary IoT cloud account. In places where filming shared hallways or common areas is legally or socially problematic, an audio-only phone-call doorbell is often a better fit. Twilio remains a third-party SIP/PSTN bridge, but the doorbell's SIP leg uses TLS signaling and SRTP media; there is no Amazon Ring/Tapo-style cloud service that can activate a camera, collect video, or change device behavior remotely without you controlling the software on the board.

License

Software, firmware, scripts and documentation are released under the MIT License unless stated otherwise. Hardware design files are released under CERN-OHL-P-2.0. See LICENSE for the full license map.