Calibration and quality tiers
Informed by ../acoustic-survey-methodology.md: professional surveys use Class 1 instruments, pre/post calibration, fixed positions, station-to-station reporting, and layered metrics. Crowdsourced phone data is survey-grade at best and must be labelled accordingly.
Calibration hierarchy
| Level | Description | Typical use |
|---|---|---|
| 0 | Unknown consumer device | Informal only |
| 1 | Known model, no project profile | Tier C candidate |
| 2 | Model profile from controlled experiments | Tier B |
| 3 | Individual field offset vs reference | Tier B or Tier A eligibility under the documented protocol |
| 4 | External calibrated microphone | Tier A candidate |
| 5 | Professional SLM survey protocol | Tier A / reference campaigns |
Quiet-room / noise-floor check may detect obstruction or malfunction; it is not calibration.
Quality tiers (accepted working labels)
| Tier | Meaning (working) |
|---|---|
| A | Reference equipment or validated external mic under documented protocol |
| B | Supported consumer model with experimental correction profile |
| C | Project app on unprofiled device; complete metadata; no major quality warnings |
| D | Imported recording (e.g. Voice Memos) or incomplete session |
| E | Manual/legacy/insufficient technical metadata |
Low tiers must not silently appear equivalent to high tiers on the public map.
Uncertainty dimensions (not one score)
Report separately where possible:
- Acoustic level confidence
- Frequency content confidence
- Journey assignment confidence
- Device calibration confidence
- User metadata confidence
- Subjective response confidence (if present)
Quality flags (non-exhaustive)
Clipping, dropout, sample-rate change, Bluetooth input, voice-processing mode, AGC suspicion, obstructed mic, pocket/bag, call/interruption, too short, missing stations, implausible speed, walking mixed with train, route mismatch, duplicate, excessive speech, unknown device profile.
Recommendation
Gate public map colouring on tier + sample size; always show tier mix and uncertainty. Store raw derived values plus optional correction factor without baking corrections irreversibly into the only stored number.
Confidence: Medium (tier boundaries need pilot data).
Depends on experiment/legal/user-test? Experiments for model profiles; pilot for tier gates.
Links: ../research/04-calibration-uncertainty.md; ../research/02-acoustic-methodology.md; ../decisions/ADR-006-quality-tiers.md; H06 — Quality tiers.