Speech models are trained on human voices. They learn phonemes, prosody, spectral envelopes, formant transitions — the acoustic features that distinguish “cat” from “bat,” question from statement, one speaker from another. These features are specific to the human vocal tract: a tube roughly 17 centimeters long, with resonances between 300 Hz and 5 kHz, producing sounds shaped by lips, tongue, and palate.
Underwater acoustic targets — ships, submarines, marine mammals — produce sounds through entirely different mechanisms. A ship's propeller cavitates, creating broadband noise with blade-rate harmonics. A whale vocalizes through laryngeal folds and nasal air sacs. An underwater explosion is an impulsive shock followed by bubble oscillation. None of these sound like speech. The frequency ranges partially overlap, but the spectral structures, temporal dynamics, and noise environments are fundamentally different.
Speech large models transfer to underwater acoustic target recognition with over 99% in-domain accuracy (arXiv:2601.18086). Models trained exclusively on human speech — with no underwater data in their pretraining — serve as feature extractors for classifying ships, submarines, and marine mammals. Fine-tune the final layers on a modest underwater dataset, and the speech-derived features outperform features trained from scratch on underwater data alone.
This should not work. The features useful for distinguishing vowels should not be useful for distinguishing ship types. But it does work, which means the speech models are learning something more general than speech — some representation of acoustic structure that transcends the specific domain of human vocalization. Spectrotemporal modulations, onset dynamics, harmonic relationships: these are properties of sound, not properties of voices.
The ear trained on words hears the ocean. The representation is deeper than the training data.