R2 — Acoustic methodology: metrics, standards, claim limits
Decision-oriented research for TUNES Wave 1. Must be read with acoustic-survey-methodology.md. TfL R3291 (FOI) is a methodology reference only — not redistributable data and not affiliation or endorsement.
Owns: which metrics to collect/publish; how standards apply (or do not); multi-metric and station-to-station discipline; what consumer devices cannot claim.
Does not own: citizen-science governance (R1); iPhone AGC/API experiment design detail (R3); full calibration hierarchy (R4); claim-language legal sign-off.
1. Decision questions
- What metric set should the iOS collector compute and what should the public map / open dataset expose by default?
- Which ISO/IEC (and related) standards apply directly, which are informative, and which claims are unsupported on consumer devices?
- How do we reuse professional survey discipline (sections, layered metrics, provenance, grading) without pretending to be Class 1?
Open questions touched: #6 (research / profiled / informal grades), #11 (exposure / discomfort / loudness / risk wording).
2. Professional discipline to reuse (from methodology notes)
From R3291 practice notes + ISO 3381 framing — transferable system, not transferable Class 1 status:
- Traceable (or best-available characterised) instrument with stated limits
- Pre/post checks where feasible; discard/suspect runs that fail stability checks
- Justified measurement position/height (or honest passenger-position metadata)
- Controlled and disclosed conditions; record exclusions
- Canonical spatial unit = station-to-station section with duration
- Layered metrics — never blend A / C / peak / bands into one unexplained number
- Full provenance: device, software/pipeline version, corrections stored separately
- Honest grade: in-service crowd data is survey-grade at best
TUNES adds passenger loads, ordinary behaviour, and optional perception — those enrich context; they do not upgrade the acoustic grade.
3. Spatial unit and analysis time
3.1 Station-to-station section
Working assumption (accepted for research direction): the canonical public comparison unit is a station-to-station section with duration T.
Why this matches professional practice:
- Maps cleanly to network topology
- Comparable across days and contributors
- Forces every Leq to carry an averaging time
Implementation notes (product detail owned later by R6/R7 + ios/web):
- Report both directions separately when the network model distinguishes them
- Exclude or flag non-representative legs (e.g. held mid-section) with disclosed reasons — same honesty pattern as professional footnotes
- Do not publish a section Leq without
T
3.2 Homogeneous sessions
Participatory noise-mapping practice (NoiseCapture-style protocol ideas): keep a session in a homogeneous environment; start a new logical segment when the environment changes (platform vs moving train). Underground GPS is unreliable — snap to station-pair geometry, don’t trust raw fixes as acoustic truth.
4. Metric set
Prefer derived features computed on device (or in a versioned pipeline) over uploading raw audio. Keep fields separate.
4.1 Core metrics (launch-directed)
| Metric | Role | Notes |
|---|---|---|
| LAeq,T | Primary continuous level for section comparison | Always with A-weighting and duration T (ISO 1996-1 framing) |
| LCeq,T | Low-frequency / flatter weighting companion | LCeq − LAeq large → significant LF energy (rule-of-thumb diagnostic, not a health claim) |
| LCpeak | Instantaneous C-weighted peak | Not a time average; useful for impulsive events; regulatory context only |
Percentiles (e.g. LA10, LA50, LA90 over T) | Distribution within the section | Helps distinguish steady rumble vs intermittent events; exact set finalized in architecture |
| One-third-octave L_eq bands | Spectral diagnosis | Prefer IEC 61260-style band centres; store as derived spectrum, not a single “tone score” |
Z-weighting / unweighted levels: useful for research exports when computationally available; not required for passenger-facing primary colouring.
4.2 Time structure
- Section-level aggregates are the public unit.
- Optional finer internal frames (e.g. short Leq blocks) may support boundary detection and QA — version them; don’t silently redefine historical section values (R10).
4.3 Psychoacoustic and event candidates (research backlog, not launch claims)
Candidates for later derived features / research layers only until validated:
| Candidate | Why interesting | Claim limit |
|---|---|---|
| Tonality / wheel-squeal indicators | Passenger-relevant tonal events | Algorithm + version required; not “official squeal index” |
| Sharpness, roughness, fluctuation strength | Timbre / modulation character | Psychoacoustic models are method-dependent; keep out of unexplained blended scores |
| Impulsiveness indicators | Door slams, bangs | Complement LCpeak; don’t replace it |
| Speech masking / conversation difficulty proxies | Links to accessibility concerns | Prefer pairing with subjective items (R9); avoid medical claims |
| Vibration-related experience | Felt vibration ≠ mic pressure | Needs sensors/methods beyond SLM metrics; separate channel |
Do not ship a single “discomfort score” that mixes objective bands and survey answers without an explicit, documented model version.
4.4 What not to do
- One decibel number as a complete passenger-experience description
- Averaging LAeq with LCpeak or octave bands into one field
- Baking calibration offsets into published values without retaining unadjusted + offset provenance
- Calling crowd results engineering-grade
5. Standards applicability
5.1 Directly informative for method design
| Standard | Relevance to TUNES |
|---|---|
| ISO 3381:2021 | Interior railway noise methods; heights 1.2 m seated / 1.6 m standing; Class 1 instrumentation requirement; grades via ISO 12001; in-service / relaxed monitoring is below engineering grade — crowd Tube runs are survey-grade at best |
| ISO 1996-1:2016 | Definitions of Leq, weightings, how to express quantities — always quote weighting + duration |
| IEC 61672-1 | SLM Class 1 vs Class 2 performance — defines what phones are not |
| IEC 61260 | Fractional-octave filters — target for band features |
| IEC 60942 | Sound calibrators — relevant when external mic + calibrator exist (Tier A paths) |
| ISO 12001 | Engineering vs survey grade language — use for honesty labels |
5.2 Occupational / environmental context (informative, not TUNES compliance)
| Instrument | Use carefully |
|---|---|
| Control of Noise at Work Regulations 2005 (UK) | Explains why dose metrics (e.g. LEX,8h) and LCpeak appear in workplace framing — does not make TUNES an occupational compliance tool |
| Environmental assessment practice | Informative for descriptor vocabulary; passenger map ≠ statutory noise map |
5.3 What standards do not authorize on consumer phones
Uncalibrated or lightly characterised iPhones cannot honestly claim:
- Conformance to IEC 61672 Class 1 (or typically even Class 2 without demonstrated verification)
- ISO 3381 engineering-grade type-test equivalence
- Substitution for operator acceptance tests or regulatory surveys
- Traceable chain to a national reference without documented calibrator + procedure
- Health risk classification, hearing-damage dose, or medical advice from map colours
Smartphone literature (e.g. NIOSH SLM app programme on iOS; peer-reviewed app-vs-SLM evaluations) supports treating apps as adjunct / screening tools with model- and condition-dependent error — cite primary sources when quoting any numeric tolerance; do not invent project-specific accuracy figures here. Any TUNES tolerance bands must come from TUNES experiments (R3/R4).
6. Measurement positions and passenger reality
ISO 3381 fixed heights (1.2 m / 1.6 m) are the professional ideal. Ordinary passengers will sit, stand, hold phones at variable heights, and sometimes obstruct mics.
Decision posture:
- Capture position metadata (sitting/standing; phone in hand vs pocket/bag; approximate carriage location) rather than pretending ISO heights.
- Protocol guidance: phone in hand, mic unobstructed; don’t tap/handle mid-run; avoid pocket/bag for higher tiers (R1/R3).
- Pocket/bag or voice-processed sessions → quality flags and tier demotion, not silent “correction to ISO.”
7. Quality tiers and grading (working)
Align acoustic honesty with quality tiers A–E (glossary working labels):
| Tier | Acoustic reading |
|---|---|
| A | Reference SLM or validated external mic + documented protocol — still label survey vs engineering per actual method |
| B | Supported iPhone model with experimental correction profile — relative and profiled absolute claims only within published uncertainty |
| C | Project app, unprofiled device, complete metadata, no major flags — useful for relative patterns with caveats |
| D | Imports / incomplete control (e.g. Voice Memos path if ever allowed) — research curiosity / weak public weight |
| E | Insufficient metadata — archive or exclude from passenger aggregates |
Crowd in-service data maxes out at survey-grade at best even at Tier A when conditions are not type-test controlled.
8. Public claim language (decision draft for ADR)
| May say (with caveats) | Must not say (foundation stage) |
|---|---|
| Section LAeq from passenger samples; sample size; tier mix | “Scientifically validated public map” without tiers + experiments |
| “Often louder in our samples than section Y” | “Unsafe,” “hearing damage,” “exceeds legal limit” |
| Low-frequency energy suggested by LCeq−LAeq | Medical diagnosis or accessibility guarantee |
| Survey-grade passenger monitoring | Engineering-grade / Class 1 equivalent / TfL-approved |
Allowed vocabulary for exposure, discomfort, loudness, and risk follows ADR-011; any broader claim requires new acoustic and legal evidence.
9. Minimum derived payload (schema starter)
Mirror auditability from professional tables without Class 1 pretence:
- Network: line, direction, from/to station (section)
- Metrics: name, weighting, value, unit, duration
T - Optional: percentiles, 1/3-octave vector, peak
- Device: model, OS, app version, pipeline version
- Calibration: offset if any (stored separately), tier, quality flags
- Context: position/height proxy, crowding, exclusions
- Grade label: survey-grade at best for crowd collection; a documented reference protocol states its actual grade separately
- Alignment: snapped vs GNSS accuracy note
Prefer this derived package for upload; raw PCM stays on device by default (R5).
10. Repo implications
| Repo | Acoustic obligations |
|---|---|
tunes | Metric definitions, claim limits, methodology versions, open releases with provenance |
tunes-ios | On-device derived feature computation, flags, section timing, offline buffering |
tunes-web | Aggregation per section, multi-metric display, uncertainty, no blended mystery scores |
Recommendation
Adopt a multi-metric, section-based, survey-grade-honest methodology: public comparison on station-to-station LAeq,T plus LCeq,T, LCpeak, percentiles, and 1/3-octave derived features; never blend metrics or objective/subjective scores; label crowd data survey-grade at best; use ISO 3381 / IEC 61672 / ISO 1996-1 as normative vocabulary and honesty bounds, not as a certificate for iPhones. Psychoacoustic indices stay in research backlog until versioned and validated. Finalize numeric uncertainty bands only after R3/R4 experiments — do not invent them.
Confidence
High for section unit, layered metrics, survey-grade labelling, and “phones ≠ Class 1.” Medium for exact percentile set and which psychoacoustic candidates enter v1 derived schema. Low for public risk/exposure wording until legal/acoustic ADR.
Depends on experiment/legal/user-test?
Yes. Device accuracy and orientation experiments (R3); calibration hierarchy (R4); user tests for how multi-metric maps are understood; legal review for claim language (#11); optional expert review of band/psychoacoustic feature specs.
Links to related docs
- ../acoustic-survey-methodology.md — required companion
- ../governance/project-charter.md
- ../02-research-map.md
- ../01-assumptions-and-open-questions.md
- H10 glossary — section, derived features, survey-grade limits, tiers
- 01-citizen-science.md
- 03-consumer-device-limits.md
- H12 measurement philosophy
- H15 railway noise taxonomy
- Downstream: R4 calibration/uncertainty; R6 journey model; claim-language ADR; open-data layers (R10)
Primary standards and references (do not treat as TUNES results)
- ISO 3381:2021 — https://www.iso.org/standard/77368.html
- ISO 1996-1:2016 — https://www.iso.org/standard/59765.html
- IEC 61672-1, IEC 61260, IEC 60942; ISO 12001
- NIOSH SLM app programme — https://www.cdc.gov/niosh/noise/about/app.html
- NoiseCapture protocol FAQ — https://noise-planet.org/faq_NoiseCapture.html
- TfL R3291 via FOI — methodology reference only; not open-licensed data; no TfL endorsement