Scaling a self-supervised speech model from 126 to 4,017 languages reveals a phase transition. Phylogenetic recovery — the ability to reconstruct language family trees from audio features alone — is stagnant through 1,000 languages. At 4,000, something qualitative happens: the model suddenly resolves both clear lineages and complex contact patterns. The improvement is not gradual. It is a threshold.
Kim, Kim, and Mortensen (arXiv:2603.07238) find that the 4,017-language model recovers a robust macro-cluster linking Papuan, Oceanic, and Australian languages — families with no established genetic relationship. These language groups have been geographically proximate for tens of thousands of years. The model detects signatures from acoustic features alone, without any textual or morphological input.
The result is either a methodological artifact or a genuinely new empirical finding about language prehistory. If artifact, the shared acoustic patterns reflect nothing deeper than geographic proximity and chance. If real, there are acoustic traces of deep contact — phonological substrate effects, shared articulatory habits — that survived millennia of divergence and are invisible to morphological comparison but detectable at sufficient scale.
The phase transition itself is the structural claim. Below a critical number of languages, the model lacks the cross-linguistic context to distinguish deep signal from noise. Above it, enough bridging languages exist to connect distant families through chains of acoustic similarity. The percolation metaphor is apt: isolated clusters of related languages suddenly connect when enough intermediate nodes are added. The discovery isn't in the model architecture. It's in the data threshold — a property of the world's linguistic landscape, not of the learning algorithm.
Kim, Kim & Mortensen, “Scaling Self-Supervised Speech Models Uncovers Deep Linguistic Relationships: Evidence from the Pacific Cluster,” arXiv:2603.07238 (March 2026).