What the engine computes
Every reading is built from the same pipeline. The first five stages are implemented and covered by unit tests, including a check of the digital A-weighting filter against the standard curve at reference frequencies. The last three are the calibration work in progress.
01
Measurement audio sessionAVAudioSession .measurement mode reduces iOS voice processing on the microphone path.
Built02
Uncompressed PCM captureHardware sample rate, DC offset removed before any weighting.
Built03
A-weighting filterPrewarped digital IIR designed per sample rate; unit-tested against the IEC 61672 curve.
Built04
LAeq 30 s and Fast peakLinear-domain energy averaging over a rolling 30-second window plus a 125 ms Fast maximum.
Built05
State detectionSettling, stable, changing, very quiet, and clipping states so a number is never shown without context.
Built06
Per-model calibration profileSigned catalog mapping iPhone model and engine version to a correction curve and uncertainty.
In progress07
Class 1 reference comparisonLab session with a traceably calibrated Class 1 meter, 35 to 100 dBA, eight sound types.
In progress08
Public dataset and reportRaw readings, error statistics, and known limitations by model, sound type, and orientation.
In progressWhy a 30-second average
An instantaneous level jumps with every rustle. A 30-second A-weighted equivalent continuous level (LAeq,30s) smooths that into something a parent can compare between two placements. Alongside it we show the recent Fast-weighted maximum so a loud transient in a sound file or a speaker does not hide inside the average.
Readings also carry a state: settling, stable, changing, very quiet, or too loud to estimate. A number without its state is not shown.
The reference comparison protocol
| Reference | A traceably calibrated Class 1 sound-level meter with a fresh acoustic calibrator check before and after each session. |
|---|---|
| Devices | Multiple physical units per model across at least eight iPhone model families, with and without common cases, in flat and upright orientations. |
| Levels | Nursery-relevant 35 to 70 dBA and event-relevant levels up to 100 dBA, stepped in 5 dB increments. |
| Sources | White, pink, and brown noise, a fan, music, crowd recordings, traffic, and impulsive event samples such as applause and fireworks. |
| Geometry | Source at 30 cm, 1 m, and 2 m from a mattress-height microphone position in a furnished room, plus a low-reflection control. |
| Statistics | Mean error, median absolute error, 95th-percentile absolute error, and worst underreading, reported by model, sound type, level band, case state, and orientation. |
| Publication | A readable report, the supported-device table, the raw dataset, the engine version it applies to, and the models that failed or were not tested. |
We publish failures. If a model’s 95th-percentile error is too wide for nursery bands, the app treats it as unvalidated and says so. The report will list it either way.
Planned device coverage
The first lab batch prioritizes current phones as Source and the phones most likely to be sitting in a drawer as Meter. Status will change to “Validated” with a linked report and a ± figure as each model completes.
| iPhone model | Role | Validation | Note |
|---|---|---|---|
| iPhone 16 Pro | Source and Meter | Pending | First lab batch |
| iPhone 15 | Source and Meter | Pending | First lab batch |
| iPhone 14 | Source and Meter | Pending | First lab batch |
| iPhone 13 | Meter | Pending | Common drawer phone |
| iPhone 12 | Meter | Pending | Common drawer phone |
| iPhone 11 | Meter | Pending | Common drawer phone |
| iPhone SE (2nd and 3rd gen) | Meter | Pending | Second lab batch |
| iPhone XR / XS | Meter | Needs units | Needs units |
Experiments we will publish alongside the report
- Distance experiment. White, pink, and brown noise at 30 cm, 1 m, and 2 m from a mattress-height microphone, phone models and common nursery machines side by side with the reference.
- Output-route experiment. The same sound and app volume through the iPhone speaker, a Bluetooth speaker, and a HomePod, showing why a saved room setup is invalidated when the route changes.
- Case and orientation experiment. How much common cases and flat versus upright placement shift readings on each model.
- Venue-position experiment. Dance floor, near-speaker table, distant table, lobby, outdoors: average, peak, and duration with stated uncertainty.
- Sound-catalog audit. Level matching, true-peak ceiling, and loop-seam results for all eight sounds.
Known limitations
- Built-in microphones vary between individual units of the same model. Published uncertainty covers the units we tested, not every phone ever made.
- Very quiet rooms may fall below a model’s reliable floor. The app reports “very quiet” instead of a number it cannot support.
- Loud, impulsive sounds can clip the microphone. The app reports “too loud to estimate” rather than an underreading.
- iOS updates can change the audio path. Each calibration profile is tied to an engine version, and a mismatch reverts the model to unvalidated.
Sources
- American Academy of Pediatrics, Preventing Excessive Noise Exposure in Infants, Children, and Adolescents (2023)
- Hazardous sound outputs of white noise devices intended for infants (2021)
- Apple, AVAudioSession measurement mode
- NIOSH Sound Level Meter app, the right general-purpose recommendation when a parent needs a workplace-style meter rather than a nursery tool.