TonalWatch

Inspiration

Virginia holds more data centers than anywhere on Earth — and when they run out of industrial land, they get built next to suburbs. Residents document a persistent low hum. Standard sound meters can't hear the problem: Loudoun County's own noise page admits its dBA standard "does not capture low frequency noise or the discomfort caused by steady tonal sounds." A resident complains; the county's meter says everything is fine; the complaint dies. The harm persists because it is invisible to the measuring tool the law uses.

We picked this problem after screening 18+ alternative data-center solutions (waste-heat reuse, rainwater capture, siting tools) and finding the same thing each time: mature infrastructure already exists. For community noise around data centers — nothing. No map, no complaint protocol, no tooling anywhere.

What it does

TonalWatch measures the exact thing the standard can't see. Any phone or laptop records a 5/15/30-second sample; the browser computes two official sound weightings at once — A-weighting (the way standard meters "hear," the way the law counts) and C-weighting (much more sensitive to low frequencies, the way your body registers a hum). The difference (C−A) isolates low-frequency content: when C reads 15+ dB higher than A, that is the tonal signature the county's number misses. Constant volume errors cancel in the difference — that's the honest part.

Users get a plain-language verdict (no tonal noise detected / tonal noise likely / sample unusable — with reasons like background noise, wind, or clipping), can compare consecutive captures, learn the science in one page written for two audiences (a neighbor who just wants clarity, a student who wants the mechanics), and log their reading on a community map of data-center sites — crowd-evidence in the style of Waze, but instrumented: each pin is a repeated physical observation, not a testimony.

How we built it

TypeScript + React (Vite), a Web Audio capture pipeline, and a deterministic DSP engine in a Web Worker: FFT with Hann windows, IEC A/C weighting curves normalized at 1 kHz, frame persistence tests for the experimental peak check.
A test-first engine: 74 unit tests including synthetic fixtures with known answers (a 63 Hz hum + harmonics reads ~+20 dB gap; a pure 1 kHz tone reads ~0 — a deliberate counterexample showing why a gap alone is not a tone detector).
A standalone field lab for hardware validation, plus quality gates (silence/noise-floor/clipping checks, honest "sample unusable" states instead of confident numbers from bad input).
Local-first map (Leaflet + OpenStreetMap) with explicit provenance labels; live demo works offline.

Challenges we ran into

The dBFS ↔ SPL trap: our first spike read −150 dB because a dead input device's raw digital floor was being displayed as if it were sound pressure. Fixed with device selection, live input meters, and honest "digital level, uncalibrated" labeling everywhere.
The calibration wall: absolute sound levels need lab calibration per device. Our answer: don't sell absolute levels — sell the difference (C−A), which cancels any constant volume offset, then validate cross-device agreement in the field.
Field reality: two team members ran a 29-capture validation campaign across campus in one afternoon — exploratory, honestly labeled, but it measured three things live (airflow from fans masks the tone; the signature reads clearer at distance; human broadband noise collapses sensitivity). Every claim in the pitch is tagged with exactly what the data supports.
Merging three humans + two AI agents on one repo mid-competition (scope decisions stayed human; agents proposed, humans disposed).

Accomplishments that we're proud of

Our field exploratory test found a gap that TRIPLED (≈6 dB → ≈20 dB) with distance from an industrial AC source while loudness fell — matching the documented complaint pattern "audible over a mile away, but compliant at the fence." Consistent with our design physics.
The industry's own noise-control literature admits the tonal hum data residents need is effectively unobtainable — we built the resident-side instrument for that gap, verified on consumer hardware.
Prince William County's May 2026 ordinance (PWC 25-62: 73 dBC day / 68 dBC night) already codifies the A-vs-C approach into law — and Loudoun's noise-code rewrite is scheduled for 2027. Community instrumented evidence is the missing layer between 5,000 frustrated complaints and the scheduled rewrite.

What we learned

Measurement is a protocol, not a gadget: stand still, stay quiet, stay out of the airflow, take three captures, keep the middle value, log the conditions. The method came from our own field data.
Honesty scales: every number on screen carries its own limit ("uncalibrated," "exploratory," "screening only") — and that made the project more persuasive, not less.
The airport asymmetry a judge handed us: nobody moved next to an airport and complains about planes. Data centers moved next to THEM. That frame changed how we pitch the whole problem.

The problem/solution, complete — and the real fix after the map

The honest objection: a complaint map doesn't make the hum quieter. True — and it's not supposed to. It's the evidence layer the whole solution chain runs on:

Residents → evidence → the legal machinery → fewer tons of noise at the source.

Document (built today): instrumented, protocol-based readings turn "I hear a hum" into data with a method attached.
Aggregate (built today): the map converts scattered complaints into a pattern the county can't average away.
Trigger (the county's own process): repeated verified screening measurements are the thing that gets a licensed acoustical survey commissioned — the step residents can't afford alone.
Regulate (already scheduled): Loudoun's noise-code rewrite lands in 2027. PWC's Ordinance 25-62 proves the mechanism exists in Virginia today: the same A-vs-C measurement logic, already written into law. Evidence from the map is exactly the input that rewrite will consume.

And past the map — actual mitigation, where our measurements become the tool:

Walls and berms — already failing, measurably. Low-frequency sound diffracts around barriers like water around a rock; Loudoun's existing berms can't stop the 50–250 Hz tonal band. Our measurements would show that before spending another dollar.
Distance and siting — the strongest controls. The signature reads clearer at distance while loudness falls; siting and setbacks are the highest-leverage fix, and instrumented data is how a county defends its setback rules.
At the source — the fix the industry itself documents. Operators already retrofit fan blades to eliminate specific tones (AWS's own noise-control literature documents exactly this), often while improving fan efficiency. Localizing a tonal peak to a specific fan speed makes a retrofit work order possible — quieter AND cheaper to run.
What gets worse: nothing at the source changes until the noise is legible. Every mitigation above runs on the same input: measured, repeated, honest data. That's the layer we built.