Soundwalks (walk mode)
A walking recording breaks the toolbox's core assumption — one microphone,
one place, one soundscape whose statistics converge with time. On a
soundwalk the non-stationarity is the signal, so walk inverts the
pipeline: segment the recording into sub-soundscape zones first, then
describe each zone as a short stationary session, the way ISO/TS 12913-2
soundwalks treat their stop points.
Usage
ambiscape analyze WALK/ # cache features first
ambiscape walk WALK/ # zones, cadence, regimes
ambiscape walk WALK/ --gps track.gpx # + distance, speed, route map
Outputs land in WALK/analysis/: walk.md (the zone table),
walk_zones.tsv (one row per zone), route_profile.png (spectrogram,
level, cadence and diffuseness strips with zone boundaries), and — with a
track — route_map.png (the route coloured by zone).
What it computes
- Zones from multivariate change-point detection on the cached 1 Hz
features (
ambiscape.segmentation): Foote novelty over log-octave powers, log-centroid, log-flatness and diffuseness, on physical dB-like scales with an absolute contrast floor (--threshold-db, default 4). Transitions on a walk are gradients; read boundaries as centres of transition zones. - Per-zone descriptors: Leq, L10−L90, band fractions, flatness, centroid, diffuseness, event rate — each zone treated as a short stationary session.
- Step cadence from the 50 Hz envelope:
steps_per_minis a walking-speed proxy,step_saliencesays how much of the zone's foreground is the walker's own feet. - A figure–ground regime per zone:
own-steps(the gait owns the foreground),external-events(things keep emerging),bed-only(nothing does),mixed. - Self-noise gating: footfalls (envelope impulses, dilated) and wind
(per-second low-band × diffuseness gusts) are masked out of a gated
level. Zones with clear gait report
leq_gated_dbfs— what the place sounds like without the walker — plusstep_time_fractionandwind_time_fraction. On the reference walk the walking group was 2–4 dB of its own quiet zones. - GPS (
--gps track.gpx): GPX track points give each zone a distance, mean speed and midpoint position; track speed is an independent cross-check on the acoustic cadence.--gps-offsetshifts session time when recorder and tracker do not share a clock (calibration.jsonclock_offset_sis applied to the session already).
Session-level honesty
analyze itself now diagnoses non-stationarity: a session whose features
hold several regimes says so in its README, because session-level
descriptors average over different places. Ecoacoustic indices (ACI,
NDSI, ADI) assume a fixed sensor and are not meaningful for a walk;
between-zone comparisons within one walk and one rig are the defensible
use of everything here.
Companions
ambiscape draft proposes keynote beds per spectral regime (so a walk
drafts one bed per sub-soundscape, not per level band) and ambiscape
taxonomy renders the Schafer timeline — the natural sequence view of a
walk — and the Schaeffer map, whose detected events inherit any named
annotation spans that cover them.
Design notes and literature grounding live in the wiki's Soundwalk-Mode page; the reference case is a 12.8-minute walk across the Blindern campus (MUS2640, 2026-08-24).