H15 — Railway noise taxonomy
This taxonomy supplies shared labels for describing major sound sources heard by passengers. It does not assert that TUNES can automatically classify them.
For: contributors, acoustic researchers, annotation designers, and pipeline implementers.
Assumptions: source labels supplement measured metrics; they do not replace LAeq,T, LCeq,T, LCpeak, or band summaries. Frequency depends on vehicle, track, speed, structure, microphone, and environment.
Detectability scale
| Term | Meaning |
|---|---|
| Potentially detectable | Derived level, spectral, peak, or motion evidence could support a candidate detector |
| Weakly distinguishable | Evidence may show an event but not identify its source reliably |
| Not established | The repository contains no validated TUNES detector or performance result |
All categories are currently not established as automatic classifiers. “Likely detectability” below describes candidate evidence only.
Source categories
| Category | Definition and characteristics | Frequency information | Likely detectability |
|---|---|---|---|
| Rolling | Continuous wheel–rail and vehicle-motion sound during travel; often a broad rumble or roar that changes with speed, track and tunnel. | Broad-band; low-frequency content may be visible in C-weighted and band summaries. No TUNES source-specific range is defined. | Potentially detectable from sustained envelope, bands and motion; source separation is unvalidated. |
| Braking | Sound associated with deceleration, including friction, traction-system or wheel–rail components. It may be broad-band, tonal, or intermittent near a stop. | No fixed range; depends on braking system and stock. | Potentially detectable from audio change plus deceleration cues; attribution is unvalidated. |
| Flange squeal | Strong tonal or narrow-band wheel–rail sound, commonly perceived as piercing and associated with curves. | Tonal location varies; TUNES has not adopted a numeric band. | A candidate tonality/wheel-squeal indicator is research backlog; event detection and naming require algorithm versioning and validation. |
| Rail joints | Short, often repeated impacts as wheels cross discontinuities or joints. | Impulsive and broad-band; no fixed range. | Peaks and periodic events may be detectable; distinguishing joints from other impacts is weak. |
| Points | Sound generated while traversing switches, crossings, or related discontinuities; may combine impacts, vibration and changing rolling texture. | Broad-band and installation-dependent. | Event presence may be detectable; automatic distinction from rail joints or vehicle impacts is weak. |
| HVAC | Heating, ventilation and air-conditioning fans, airflow, compressors, or auxiliaries inside the vehicle. Usually steady or slowly varying. | Often broad-band with possible tonal components; no TUNES range is defined. | Potentially detectable during dwell or steady conditions, but separation from traction and ambient noise is unvalidated. |
| Doors | Door movement, latches, slams, warning tones and chimes around station dwell. | Mechanical events are impulsive; warnings may be tonal. Exact bands vary by stock. | Door-energy bursts are candidate alignment hints; identifying a door event is not yet a validated classifier. |
| Passenger | Speech, movement, luggage, personal devices and other sounds generated by occupants. Occupancy is context, not an acoustic metric. | Speech and movement span multiple bands; no source range is stored by default. | Speech presence may support an excessive_speech quality flag; speaker identity or transcription is out of scope. |
| Announcements | Public-address speech, alert tones and service messages from train or station systems. | System- and stock-dependent; speech plus possible tonal cues. | Presence may be detectable. Station-name recognition is deferred, privacy-sensitive research and must not create a speech archive. |
| Construction | Maintenance, engineering, demolition, drilling or other works audible from railway infrastructure or adjacent sites. | Tool- and environment-dependent; may be tonal, impulsive or broad-band. | Weakly distinguishable without external context; no TUNES detector is specified. |
| Unknown | Audible content that cannot be assigned confidently, or a mixture of sources. | Preserve measured metrics without forcing a source range. | Required fallback whenever evidence does not justify a label. |
Annotation rules
- Preserve source label, annotator type (
user,campaign, or algorithm), algorithm/version where applicable, time scope, and confidence separately. - Permit multiple labels for mixed events.
- Use unknown rather than forced attribution.
- Do not infer operator fault, maintenance condition, health risk, or regulatory breach from a source label.
- Keep passenger perception labels such as “piercing” or “rumbling” separate from physical source attribution.
- Never replace the underlying objective metrics with taxonomy counts.
Future work
TUNES has not fixed source-specific frequency ranges, annotation enums, detector algorithms, training data, or validation thresholds. These require acoustic review, privacy assessment, labelled experiments, and a schema/ADR before public classification claims.
Related Documents
Measurement philosophy · Schema · Acoustic methodology · Journey alignment · Subjective experience · Consumer device limits · Claim language