Event timbre templates

A soundscape repeats itself: the same door, the same cup, the same bird call recur through a session. ambiscape timbre finds those recurring event classes without any machine learning. Every transient gets a spectral fingerprint — the strike-triggered rise spectrum (what appeared) plus a per-band decay slope (how it faded) — and the fingerprints are clustered by correlation distance into transparent, corpus-comparable template classes. It complements the PANNs sound-event tagger ([ml]) with a fully inspectable, no-model alternative.

ambiscape timbre <session-folder>   # needs a prior analyze run

Uses the cached spectral events to place onsets, then reads short windows of audio around each onset to build the fingerprints. Writes timbre.json (n_events_fingerprinted, n_classes, n_unclustered, and a classes list) and timbre.png (the fingerprints as a heatmap, grouped by class).

Reading a class

Each entry in classes describes one recurring sound:

• n — how many events fell into the class;
• exemplar_t0_s — onset times of up to five members, to go and listen;
• centroid_hz — the fingerprint's spectral centroid, a rough "how bright";
• decay_median_db_s — the median decay slope over the bands the event actually excited: sharply negative for a plucked or struck sound, near zero for a sustained one.

Classes are sorted by count. Singletons that match nothing are left unclustered (n_unclustered) rather than forced into a class.

In Python

from ambiscape import timbre

rise, slope, centers = timbre.event_fingerprint(take, t_onset=123.4)
labels = timbre.cluster_events(fingerprints)   # -1 = unclustered singleton

Clustering is average-linkage on correlation distance with a 0.35 threshold; two events are "the same sound again" when their rise spectra correlate tightly. Because there is no trained model, the classes mean exactly what the fingerprints say and compare directly across sessions.