Speech AI trained with no labels ends up modeling the mouth and throat, not any particular language's sounds.

This explores why speech models trained on raw audio end up modeling the physical machinery of speech (tongue, lips, jaw, vocal folds) instead of a catalog of language-specific sounds. The direct evidence in the corpus is narrow. It rests mainly on one finding: self-supervised speech models infer the articulatory processes that generate acoustics, and this applies across all human speech Do speech models learn language-specific sounds or universal physics?. The likely reason is simple. When a model learns by predicting hidden pieces of audio, the cheapest way to explain the sound is to model what produced it. Every language is made by the same kind of vocal tract, so the physics is the one thing all the training data has in common. Phonetic categories like 'the English /r/' are a layer on top of that physics, and they change from one language to the next.

The payoff is practical as well as interesting. This physical representation explains why speech models transfer so well to languages they barely saw in training. It also predicts how well a model will do on downstream tasks better than tests that check whether the model has learned phonetic categories Do speech models learn language-specific sounds or universal physics?. That carries a broader lesson about how interpretability tests can mislead. Finding a concept you expected inside a model doesn't mean the model relies on it. In language models, information that a probe can detect often has no causal effect on what the model actually generates Do language models actually use their encoded knowledge?. If you probe a speech model for phonemes because that's how linguists carve up speech, you might find them and still miss what the model really runs on.

The most useful comparison is with text. One line of argument holds that text-only LLMs are like the prisoners in Plato's cave. Writing is a lossy human abstraction that strips out the physics, geometry and causality of the world that produced it, so text models handle symbols with no access to where they came from Are text-only language models fundamentally limited by abstraction?. Audio hasn't been through that compression. The waveform still carries traces of the body that made it, so a speech model gets the grounding a text model can't. Put the two findings side by side and they suggest a general rule: a model learns the deepest causal structure its input still contains. For text, that structure is linguistic convention. For audio, it is physiology.

A weaker supporting thread is that neural networks often organize what they learn into clean, structured geometry without being told to. For example, language models encode grammatical relationships as distances and angles between embeddings How do language models encode syntactic relations geometrically?. Speech models settling on an articulatory 'coordinate system' fits that pattern, though nothing in the corpus tests the link directly.

A caveat: the corpus has one strong source on this question. It doesn't yet cover how these articulatory representations develop during training, which architectures produce them, or how they relate to speech synthesis. The comparison with text is the most transferable idea here. What a model can learn about the world is limited by how much of the world survives in its training data.