Can an AI that learns speech just by listening also figure out how our mouths physically make sound?

This explores whether speech models trained without phonetic labels pick up the physical mechanics of how voices are made, or just the sound inventories of the languages they hear. The corpus's most direct answer is yes, and it's a stronger yes than you might expect. Self-supervised speech models seem to infer the causal articulatory process itself: how the tongue, lips, and vocal tract shape airflow into acoustics. They don't stop at language-specific categories like 'the English /r/' Do speech models learn language-specific sounds or universal physics?. Because those physics are the same for every human speaker, this explains why these models transfer so well across languages. The less obvious finding is that how well a model captures articulation predicts its downstream performance better than the standard practice of probing it for phonetic labels. So the labels researchers used to treat as the gold standard may be measuring the wrong thing.

This becomes more interesting next to a parallel argument about text models. One line of thinking holds that text-only LLMs are stuck in Plato's cave: writing is a lossy human abstraction that has already stripped out the physics, geometry, and causality of the world, so models trained on it fail in predictable ways at physical reasoning Are text-only language models fundamentally limited by abstraction?. Speech models are a useful counterexample. Raw audio hasn't yet been compressed into symbols, so the physical process that produced it is still in the signal, and a model trained to predict that signal can recover it. Taken together, these suggest that what a model can learn depends heavily on how much of the generating process survives in its training data.

A third note makes a related point from the opposite direction. LLMs can be read as running Saussure's idea of language as a purely relational system: they learn meaning from how words relate to other words, with no anchor in the world Can language models learn meaning without engaging the world?. Speech SSL models suggest the opposite can happen. When the input carries traces of a physical source, unsupervised learning reaches past the relationships between units to the mechanism underneath them.

There's also a general pattern of networks spontaneously building structured internal geometry they were never told to build. Work on text models finds that syntax is encoded in something like polar coordinates inside activations How do language models encode syntactic relations geometrically?. The articulatory finding fits that pattern: given the right signal and objective, models organize their internals around the real causal structure of the data. A practical warning comes with it. If you probe only for the categories you expected, such as phonemes or parts of speech, you can underestimate or misread what the model actually learned.

A candid limit: the collection has one note squarely on this question. The surrounding material is lateral framing, not further speech-physics evidence. If you want to go deeper, start with the articulatory note, then read the Plato's-cave note as its mirror image.