Listening
Understand the science. Stand up a software loop that can ingest sperm whale audio and identify coda types from public corpora.
Whail is a wearable communication device for one of the most acoustically expressive species on the planet. We are three engineers, building it on Sundays, for one year. We have never done this before — and we are documenting every step, the boring ones included.
Sperm whales communicate in short bursts of clicks called codas. Recent science has shown these aren’t random — there is a 156-type alphabet, with structure that looks suspiciously linguistic. The inter-click intervals carry the meaning.
WHAIL is our attempt to build a small wearable that can listen to a sperm whale, recognize the coda it is producing, and — eventually, responsibly, offshore — respond.
We are unlikely to succeed in the way that phrase suggests. We are very likely to learn things by trying.
Build the thing. Then build a story about having built it. Then build the next thing.
Understand the science. Stand up a software loop that can ingest sperm whale audio and identify coda types from public corpora.
Wearable form factor. Hydrophone selection. Embedded compute. Port the classifier to hardware that can survive a brackish afternoon.
Coastal and near-shore trials. Real environments, real noise. The bench-to-ocean reality check that makes or breaks the project.
An offshore playback-and-listen experiment in the canyons off NJ/NY, under advisement of a marine acoustician. Then an honest film about what happened.
Two views of the same plan. First — the four-act timeline with the deliverable that defines each. Second — the end-to-end system, layer by layer, showing which slice ships in V1.
Understand sperm whale codas. Build a software loop that ingests audio and labels coda types from public corpora.
Wearable v0. Hydrophone selection. Embedded compute target. Port the V1 classifier to hardware. Power, waterproofing, recovery.
Coastal trials, then near-shore. Real noise floors. Calibration. Add the synthesis layer for safe, supervised playback.
Offshore canyons off NJ/NY, with a marine acoustician advising. Playback-and-listen, log everything, hold to PEPP-aligned ethics.
Where the click stream comes from. V1 uses public, published recordings — we are not collecting our own whale audio yet.
Turn raw audio into structured representations a model can use. The science says inter-click intervals carry the meaning, not the clicks themselves.
Given a coda, which of the 156 published types is it? V1 ends here — that's our minimum end-to-end loop.
Beyond "which coda" — clan, context, vowel-like spectral structure observed in 2025 findings.
Produce a coda we could (responsibly) play back. We use existing public model weights — we are not training generation from scratch.
Offshore canyons off NJ/NY. Playback-and-listen with marine acoustician advising. Not "call a whale to shore".
The minimum loop that touches every layer of the eventual system in miniature. Once this exists end-to-end, every future Sunday improves a working thing instead of staring at a blank one.