Measurement becomes a model
We publish different numbers and it's important to understand their difference. A walk produces a measured sound level in dB(A). During those walks, we record a wide number of frequencies, and from this data we create a model, focusing on lower frequencies, which we call the Rumble score.
Last updated 21 August 2026. Every figure below comes from the model run of that date.
Two metrics, many more to come
A sound level, in dB(A)
This is the literal measurement we took, weighted to correspond to frequencies that humans are most sensitive to.
A Rumble score, 0 to 100
From the measurements, we create a score, focusing on lower frequencies, think of this as the rumble from cars or subways.
Why use a different range for the model?
It simply turns out that this low range can be much better predicted, and our goal is to give you a good prediction over a model that's confidently wrong.
Value is the share of the error the model removes compared to a naive baseline (predict the mean everywhere).
| What the model was asked to predict | Held-out skill | Held-out R² |
|---|---|---|
| 50 to 250 Hz, the Rumble band | +21.7% | +0.260 |
| dB(A), what the walks report | +12.4% | +0.188 |
| 500 Hz to 8 kHz | +3.7% | +0.043 |
| Best single band, 100 Hz | +22.2% | +0.208 |
Scored against a naive average on one frozen split of walks: 6,410 windows from 78 walks to fit, 1,642 windows from 28 walks the model never saw to score. Source: data/analysis/noise/frequency_response.md, generated by scripts/analyze_frequency_response.py on 25 August 2026.
Notes and ceaveats
- We calibrated via a phone reference, not (yet) a certified meter. That is enough to rank blocks against each other, but possibly off by a few DB, and we're working on calibrating this precisely.
- EWe read on the sidewalk This is not an indoor measurement, and not meant to establish nuanced location effects.
- The honk count is a model estimate. We created a proprietary model to identify honks and have hand-reviewed a sample of the results.
The model, in numbers
- Model
- LightGBM gradient-boosted trees
- Rows it learned from
- 9,550 thirty-second windows, from 119 walks
- Date range
- 21 June 2026 to 29 August 2026
- What it predicts
- The background level (L90) of the 50 to 250 Hz band in a thirty-second window, published as a percentile rank across every modelled NYC street
- Features that carry it
-
- Hour of day 9%
- Distance to rail at street level 7%
- Distance to the nearest expressway 7%
- Distance to elevated rail 6%
- Distance to the nearest arterial road 5%
- Distance to the nearest bar or club 5%
Book a walk · $100 What we measured Open the map Honkz Pro →