Methodology

dBZ and dBC only as indicative — LF microphone rolloff uncorrected

The dBC and dBZ values computed by Noisecatcher apply the correct weighting formulae to the FFT bins, but do not compensate for the microphone's own frequency response below ~100 Hz. Consumer MEMS microphones roll off sharply at low frequencies — some attenuate 20 Hz signals by 15–25 dB compared to 1 kHz. The dBC − dBA difference is therefore informative as a relative indicator of LF content, but absolute dBC/dBZ readings should not be treated as calibrated measurements without a device-specific compensation curve.

Device microphone frequency response not corrected

A-weighting and the calibration offset together bring the mid-frequency range (200 Hz–8 kHz) close to true dB(A) for most devices. Very low and very high frequencies remain device-dependent. Readings are internally consistent within a single device and session; they are not comparable across different phone models without cross-calibration against a certified Class 1 or Class 2 sound level meter (IEC 61672-1:2013 — note: older ANSI terminology used ‘Type 1/2’; IEC 61672 replaced this with ‘Class 1/2’).

Shared map is P2P, not persistent — data remains device-local

Every pin, report, and voice note lives in your browser's IndexedDB. The P2P community layer (Gun.js) lets you share pins with nearby peers in real time, but there is no central server or shared database — Gun relays messages and peers cache what they see. If no peer is online when you publish, your pin is not saved elsewhere. The NoiseCapture GeoJSON export is a file you send manually. Collective evidence-building across time still requires coordination outside the app.

Chain of custody: capture-time hash, but no hardware attestation

Noisecatcher computes a SHA-256 hash of the session data (Leq, peak, L10/50/90, timestamp, GPS, calibration offset) before saving — not at export time. This is better than an export-time hash, but still does not constitute a forensic chain of custody under ISO/IEC 27037: a hash stored in the browser's IndexedDB can be replaced. Forensic-grade evidence additionally requires a hardware-attested timestamp at the moment of capture, an unbroken custody log, and ideally a trusted-timestamp server or immutable storage anchor. The downloadable evidence proof JSON is a useful starting document for formal complaints, not a substitute for certified forensic process.

Longitudinal visualisation: Leq trend, no Lnight/Lday breakdown

Noisecatcher now shows a Leq time-series chart across all sessions and within each Field Notebook, with ACI and NDSI sparklines. The WHO Health Risk panel (map page) computes Lnight, Lday, Levening, and Lden from all pins and session reports, with 7-day and 30-day Lnight gauges. What the Leq time-series chart still does not provide is a longitudinal Lnight vs. Lday trend — comparing how night-time and day-time levels evolve across sessions over weeks or months. That requires timestamp-aware aggregation across sessions, which the chart does not yet perform.

Distributed witnessing: shared sessions work, hardware sync does not

Multiple phones can now join a shared session via a 6-character code. Live dB readings sync across devices in real time over Gun.js P2P. TDOA (Time Difference of Arrival) triangulation estimates the distance difference and, with 3+ GPS-enabled devices detecting a shared peak, the estimated source position. The remaining limitation is clock accuracy: TDOA precision depends on the arrival timestamps, which are computed from the JavaScript event loop — not a hardware interrupt. Internet latency between peers (typically 20–200 ms) limits triangulation accuracy to tens of metres at best. Hardware-synchronised clocks (GPS-disciplined, IEEE 1588 PTP) would be needed for sub-metre accuracy.

YAMNet classification: training-data bias

The YAMNet classifier was trained on AudioSet — YouTube audio predominantly from wealthy, urban, anglophone environments. It displays results down to 2% confidence with no 'unknown' fallback. Misclassification in non-Western acoustic contexts is documented but not surfaced to the user. Treat classification results as starting hypotheses, not verdicts. Your own knowledge of local acoustic culture takes precedence.

NDSI: consumer microphone rolloff affects biophony bands

The NDSI biophony band (2–8 kHz) is within the range most consumer microphones handle well. The anthrophony band (200 Hz–2 kHz) includes some low frequencies where microphone response may vary. The index is useful as a relative, comparative measure — comparing the same location at different times, or different locations on the same device — more than as an absolute ecological score.

Psychoacoustic metrics: JavaScript approximations of ISO 532-1

Loudness, Sharpness, Roughness, and Psychoacoustic Annoyance follow the Zwicker model implemented in JavaScript from the FFT data. They are indicative, not certified. A respondent's psychoacoustician working from calibrated studio recordings and validated software will produce more reliable values. Use these metrics for awareness and pattern documentation, not as the sole quantitative basis for formal proceedings.

Long-form content is English-only

UI strings are translated across 16 locales. The Abécédaire, About, Methodology, and Act pages are hardcoded English JSX. The political ambition of the tool is global; its primary analytical language is not.