This page records how the prototype is actually built, the decisions behind it, and the limits of what has been measured. It is a development system, not a certified automotive product.
Roof node · audio capture
The design does not divide the array across two controllers. One ESP32-S3 captures all four channels using the left/right slot selection built into the microphones and two I²S peripherals sharing one clock domain.
North, south, east and west microphones are positioned on a 190 mm cross. Each opposing pair shares an I²S data line: one device drives the left slot and the other drives the right slot.
| East / West | I²S0 master · DATA GPIO 4 |
|---|---|
| North / South | I²S1 slave · DATA GPIO 14 |
| Clock | BCLK GPIO 6 · WS GPIO 5 |
| Sample format | 44.1 kHz · 32-bit stereo slots · 24-bit microphone data |
| DSP input | Top 16 bits used by the correlation pipeline |
I²S0 generates BCLK and WS. The ESP32-S3 GPIO matrix routes those signals internally into I²S1, so the second pair runs as a slave without separate physical clock pins. Startup loopback and calibration checks verify that both buses are active and aligned.
Classification and bearing
The TinyML classifier decides whether the window contains a siren. Direction finding then uses time differences across the array. That separation keeps the bearing pipeline focused on relevant audio.
The roof-node firmware uses an Edge Impulse audio export with MFE features. Training data includes real roadside ambulance recordings captured with traffic, wind and Doppler shifts rather than relying only on clean online samples.
Time-delay estimates from east/west and north/south pairs form the two axes used by atan2. With a 190 mm baseline at 44.1 kHz, one sample is approximately 2.3° near broadside. Eight-times interpolation gives a theoretical delay-grid step near 0.29°.
atan2. The design therefore does not claim that temperature compensation improves bearing accuracy.Power integrity
The roof node has a radio, switching converter and sensitive digital microphones in the same enclosure. Battery protection, conversion and filtering were designed together instead of treated as an afterthought.
Cabin node · Copilot
The cabin architecture separates time-sensitive safety I/O from the richer user interface. An ESP32-S3 handles embedded sensors and links; a Raspberry Pi runs the display, cameras and higher-level application.
Emergency-vehicle transponder
The transponder is a separate demonstrator intended for ambulances and other emergency vehicles. It broadcasts a compact position message that nearby SirenSense prototypes can place on their map.
| Controller | ESP32-S3 |
|---|---|
| Radio | SX1278 · 433 MHz LoRa |
| Position | BN-880 GNSS receiver |
| Local UI | OLED status display |
| Packet | AES-128 protected payload |
| Access | ALOHA-style jitter · 800 to 1500 ms |
A perfboard demonstration unit has been built and a PCB has been designed. The radio range has not yet been characterised in a controlled urban field test.
Validation register
Clear status is an engineering feature. These labels describe the project today and will change only when the corresponding tests are complete.
| Subsystem | Status |
|---|---|
| Roof-node PCB and enclosure | Built and demonstrated |
| Four-channel synchronized capture | Implemented with startup checks |
| TinyML siren classification | Working prototype; dataset expansion ongoing |
| GCC-PHAT direction pipeline | Implemented; field accuracy not yet characterised as a single guaranteed number |
| Cabin PCB | Fabricated; final assembly and complete bring-up ongoing |
| Raspberry Pi interface | Working software; Pi 5 and in-car validation pending |
| Transponder | Perfboard demo built; production PCB and range tests pending |
| Weather resistance | Design intent only; no IP certification |
| V16 Pro | Experimental feature; not DGT certified |
Project files