Machine listening ([ml] extra)

Two assistants, both deliberately positioned as helpers around the human ear rather than replacements for it. Both operate on the W (omni) channel, downmixed and resampled — the models are trained on 16/32 kHz mono internet audio, and quiet domestic material is out of their training distribution.

AudioSet tagging (PANNs)

ambiscape draft runs PANNs CNN14 (527 AudioSet classes) on 10-second windows around detected events and steady states, and writes the top classes into the draft's listening hints:

{"t": "07:56:16", "az": 150.0, "el": 5.0,
 "tags": [{"label": "Cupboard open or close", "p": 0.49},
          {"label": "Door", "p": 0.49}]}

AudioSet's taxonomy includes remarkably apt classes for indoor soundscape work (Air conditioning, Refrigerator, Church bell, Pigeon/dove, Water tap, Footsteps, Speech). The intended reading: the tagger says what, the intensity vector says from where — together, a labeled spatial event. Confirm by ear; scores on low-SNR ambience are suggestions.

Speech privacy gate (silero-vad)

ambiscape speechgate segment.wav              # one file
ambiscape speechgate segments/ --threshold 0.01

Reports the fraction of speech per file and a PASS/FAIL verdict (default: fail above 1 % speech). Run it on every excerpt before publishing (Freesound, Zenodo, supplementary material) — recordings made in homes routinely catch a few words the recordist forgot. Exit code 2 on any failure, so it slots into scripts.

The gate detects voice activity, not intelligibility — a conservative proxy. For a stricter check on borderline files, listen to the flagged regions (first_speech_at_s is reported).