Field Recording Protocol

The two-minute discipline that makes a session analyzable without detective work. Distilled from sessions where each item below was, at some point, learned the hard way.

Equipment & settings (keep constant)

• Zoom H3-VR (or comparable FOA recorder), AmbiX record mode, Upright position, correction filter off, 24-bit / 48 kHz.
• Note which B-format convention the recorder writes (AmbiX vs FuMa) — ambiscape auto-detects from the BWF zTRK tags, but downstream tools may not. A convention mixup silently wrecks all directional analysis.
• Check the recorder clock against your phone monthly. Clock drift of tens of seconds is invisible until you try to align with the outside world (train timetables, church clocks, other sensors).

At every setup (~2 minutes)

1. Orient — tape arrow on the recorder marks X+ (front); note its compass bearing (phone flat, aligned with the arrow).
2. Photograph the setup from above — doubles as a position record.
3. Clock sync — photograph the recorder screen next to the phone clock, or speak the phone time into the first seconds.
4. Calibration claps — snap or clap once from 2–3 known positions (window, door, dominant machine) in a noted order: direction ground truth inside the recording, immune to lost notes.
5. Room impulse — one balloon pop (or loud clap) ≥ 2 m from the mic with 5 s of stillness before and after → clean T60 via ambiscape.analysis.decay_time. Movement immediately after a clap truncates the usable decay to a fraction of a second.

Session notes (5 lines, in the session folder)

place:    <city, building, room>
position: <mic location; height; X+ bearing ° magnetic>
time:     <timezone; recorder-vs-phone offset s>
state:    <windows; HVAC/machines on-off + any toggles you make>
context:  <weather; who is present; expected sources>

Stage transitions deliberately

Toggling a machine (ventilation off for ten minutes, fridge unplugged for an hour, window opened) is the cheapest acoustic experiment there is: the transition separates source from room, reveals what the steady state masks, and gives every analysis a labeled event. Note the exact time of every toggle.

If absolute SPL matters

One SPL-meter reading of a steady source (running HVAC is perfect) next to the mic, noted with the session → a calibration.json that converts the whole session from dBFS to dB SPL.

The ISO 12913 self-report (structured)

One minute per session gives it a genuine ISO/TS 12913-2 Method-A perceptual data point — the layer no instrument can supply. Capture and storage:

When. Immediately after the session (or after each state, if you staged a transition — a drone-on and a drone-off rating from the same night is a measured-vs-perceived experiment in itself). Note the timing either way; perception drifts within minutes of leaving a room.

How. Two capture routes, both ending in the same file:

1. Spoken, in-band — in the recording's final 30 seconds, read the ratings aloud ("pleasant four, chaotic one, …"). Self-timestamping and loss-proof, like the calibration claps; transcribe when analyzing. Keep it in the last 30 s so published excerpts trim it (the speech gate will flag it otherwise — by design).
2. Phone template — a saved note skeleton you fill in and paste later.

Storage: one selfreport.json per session folder (same pattern as calibration.json and annotations.json):

{
  "instrument": "ISO/TS 12913-2 Method A, self-report",
  "reports": [
    {
      "when": "2026-07-16T08:30:00+02:00",
      "timing": "post-session",
      "state": "after pump switch-off",
      "paq": {
        "pleasant": 4, "chaotic": 1, "vibrant": 2, "uneventful": 4,
        "calm": 5, "annoying": 2, "eventful": 1, "monotonous": 3
      },
      "source_dominance": {
        "traffic_noise": 2, "other_noise": 3,
        "human_sounds": 1, "natural_sounds": 2
      },
      "overall_quality": 4,
      "appropriateness": 4,
      "note": "air pump dominant until 07:54; doves after"
    }
  ]
}

All scales 1–5 (PAQ: 1 = strongly disagree … 5 = strongly agree; dominance: 1 = not at all … 5 = dominates completely; quality: 1 = very bad … 5 = very good). From the eight PAQ values the ISO 12913-3 circumplex coordinates follow:

pleasantness = ((p−a) + cos45°·(ca−ch) + cos45°·(v−m)) / (4 + √32)
eventfulness = ((e−u) + cos45°·(ch−ca) + cos45°·(v−m)) / (4 + √32)

both in −1…+1 — two numbers per session that sit next to Leq, ψ, and R in the corpus catalog, which is exactly where measured-vs-perceived analysis starts.

Caveats to record once, not per session: this is a single-rater self-report, not the multi-participant survey the standard envisions — scientifically honest as "12913-informed self-report", and consistent rater identity is itself a longitudinal asset (one perceiver, many rooms — the StillStanding model).