cetaceanspeak

What it is. A pipeline that translates a raw WAV file of cetacean vocalization into IMASM (the categorical assembly of the Imscribing Grammar) and ranks it against human expression archetypes.

What it does. Detects acoustic units (onset, pitch, spectral centroid), compiles them into an IMASM instruction stream, measures Frobenius closure, and reports the nearest human expression type by structural distance. A 38-second humpback recording yields 125 acoustic units, closure ratio 1.0, and a nearest match of song.

Why it matters. It is direct evidence that cetacean vocalization runs the same eight-step Frobenius loop that governs written, spoken, and sung human communication: the structure is shared across species, not merely analogous.

How to use it.

uv pip install librosa soundfile numpy
uv run whale_audio.py <file.wav>
uv run whale_audio.py <file.wav> 0.04   # lower onset_delta = more onsets

---

How it works

Cetacean vocalizations share the eight-step Frobenius loop:

ISCRIB → AREV → FSPLIT → AFWD → FFUSE → CLINK → IFIX → ISCRIB

whale_engine.py compiles acoustic token sequences into IMASM and measures structural distance to six human expression archetypes. whale_audio.py drives the engine from a WAV file via librosa onset detection, pyin pitch extraction, and spectral centroid analysis.

Sample output:

── Audio: 55113001.wav  (38.4s, 14900 Hz) ──
   125 acoustic units detected
── Frobenius ──
   closure_ratio: 1.0000   paradox_count: 40   entropy_delta: 0.0000 nats
── Translation ──
   song       d=65.9492   Song structure, 4 Frobenius cycles with VINIT/TANCH bookends
   narrative  d=77.1187   Narrative, two Frobenius cycles
   question   d=84.1518   Rising-intonation question

Acoustic token labels

label	meaning
init	phrase onset
anc	phrase anchor (trailing silence)
up / dn	pitch rise / fall over threshold
link	sustained, pitch stable
rep	pitch + duration match a recent unit
fix	recurrent motif (≥ sig_repeat appearances)
split / fuse	spectral centroid bifurcates / reconverges
evalt	harmonic-rich, high SNR (social contact)
evalf	sudden energy spike (alarm)
paradox	two simultaneous fundamental frequencies

Parameters

All thresholds live in ClassifierParams:

from whale_audio import ClassifierParams, translate_wav
p = ClassifierParams(onset_delta=0.04, pitch_delta_hz=50.0,
                     anchor_silence_ms=300.0, sig_repeat=3, snr_ok_db=10.0)
translate_wav("recording.wav", params=p)

Data

Put WAV files in data/ (gitignored). Free public-domain recordings: the Watkins Marine Mammal Sound Database.

Structural type

⟨𐑦𐑥𐑾𐑿𐑞𐑧𐑲𐑠⊙𐑖𐑳𐑭⟩  O∞ tier