audio-feature

Audio-feature extraction needs signal processing to operate now — it relies on DSP techniques (FFT, filtering, windowing) to extract quantifiable properties from signals. Remove signal processing and audio-feature extraction stops working. The removal test (Law 8) passes.

Audio-feature IS a specific kind of feature — a quantifiable property from audio signals for analysis. A competent speaker calls it 'a type of feature'. Files against the nearest kind per Law 9.

Spectral-feature IS a specific kind of audio-feature — it extracts characteristics from the frequency-domain representation of audio (energy, slope, centroid). A competent speaker would call a spectral feature 'an audio feature.' Files against the nearest kind, establishing the rung for the ladder (Law 11e).

Zero-crossing-rate is a specific kind of audio feature — a numerical descriptor computed from audio signals. It IS an audio feature, not something that serves the general category.

Voice activity detection operates by extracting and analyzing audio features (spectral energy, zero-crossing rate, spectral centroids, etc.) to determine speech presence. Remove audio-feature as a concept and VAD has no features to analyze — it cannot operate. The dependency is constitutive: VAD needs audio features to function, not merely to be described.

Temporal centroid is a specific kind of audio feature — a scalar descriptor extracted from audio signals to characterize temporal distribution.

RMS is a scalar descriptor extracted from the waveform — the same kind of object as zero-crossing-rate and spectral-energy (both filed INSTANCE_OF audio-feature and accepted). It fills the energy/amplitude rung of the feature family that spectral-energy covers from the frequency side.

RT60 is a scalar descriptor extracted from an audio signal (the room's impulse response) — the same kind of object as zero-crossing-rate and spectral-energy, both accepted INSTANCE_OF audio-feature. It fills the rung of the feature family that characterizes the room rather than the source. Nearest existing kind: audio-feature (the 'measurement' entry's accepted definition carves measurement error, so it is not the target sense).

Cepstral peak prominence is a specific kind of audio-feature: a per-frame scalar descriptor extracted from a speech (audio) signal via cepstral computation, used for voice-quality analysis and classification — it fits audio-feature's accepted definition exactly (numeric descriptor derived from a digital audio signal through computational processing, serving downstream classification tasks). Nearest existing kind is audio-feature: no cepstral-feature or speech-feature entry exists. This completes the ladder cpp -> audio-feature -> feature, where audio-feature INSTANCE_OF feature is already ACCEPTED (Law 11e).