nx_bioacoustic_bench.nx

nx_bioacoustic_bench.nx -- THE RULER for ANIMAL-VOCALISATION REALISM. Operator 2026-07-25: "get nishi audio state of the art so we can generate animal sounds and a dog toy ball that makes the noise when compressed by a dog's mouth". Rung R0. BEFORE any synthesis claim we need a MEASURED NUMBER, because the only animal-sound organ that exists today (nx_audio_bark.nx) says in its own header "First-iteration timbre -- tune by ear". Tuned by ear is not a measurement, and a lane with no ruler cannot honestly say SOTA. This organ is that ruler. WHY THESE AXES. A dog ignores TV animal sound because the signal fails on several acoustic axes at once. Each axis below is one of those failure modes turned into a number that a real recording passes and a synthetic/looped/broadcast one fails: AX_BAND container bandwidth vs the DOG hearing band (dogs hear to ~60 kHz; 48 kHz media carries 24 kHz -- so a 48k file is PHYSICALLY INCAPABLE of >24 kHz no matter how good the synth). AX_ULTRA fraction of energy above 20 kHz -- content a human cannot hear and a dog can. AX_F0DYN pitch-contour movement (stddev of F0). Real calls sweep; a beep does not. AX_HNR harmonic-to-noise proxy (normalised autocorrelation peak). AX_NONLIN NONLINEAR PHENOMENA -- subharmonics, biphonation, deterministic chaos. This is the hallmark of real mammalian/avian vocal production (the source self-oscillates and jumps between regimes) and it is exactly what sample playback and hand-tuned oscillator banks do not reproduce. Highest weight. AX_NONREP non-repetition across successive segments. A looped sample cross-correlates ~1.0 with itself; a real animal never repeats exactly. This is the HABITUATION axis. AX_ATTACK envelope attack sharpness (transient realism). AX_CREST crest factor = peak/RMS. Broadcast compression crushes this. AX_RESON count of resonances (vocal-tract poles) in the long-term average spectrum. AX_FLUX spectral micro-dynamics frame to frame. A static tone has none. HONESTY DISCIPLINE. Every axis is reference-free and INTRINSIC -- it is a property of the signal itself, so no reference corpus is needed to run it and no axis can be satisfied by simply copying a reference. The ruler is only trustworthy if it REFUSES bad input, so nx_bioacoustic_bench_gate feeds it negative controls (pure sine, white noise, an exactly-looped sample, a band-limited 48k signal) and asserts each one scores LOW on the axes it should fail. A ruler that passes everything is a rubber stamp, not an instrument. See [[feedback-cynical-instruments-author-optimism-is-a-bug]]. ANTI-GAMING NOTE ON AX_NONLIN. "Loud but aperiodic" alone is NOT evidence of nonlinear vocal production -- white noise is loud and aperiodic. Real nonlinear phenomena are REGIME TRANSITIONS, so a chaotic frame only counts when the same utterance also contains periodic frames. White noise has no periodic frames, so it scores 0 here. That is the biology, and it is what makes the axis honest.