Models know when words rhyme. Is that all in the embeddings? It seems like it can't be, because there is some level of context-specific pronounciation. So where is rhyming? And how is it structured? Is there a subspace for every IPA sound?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-where-is-rhyming-2026,
author = {Holtzman, Ari},
title = {Where is rhyming?},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/2jp9uTGQmqhaV22saFW5}
}Please sign in to comment on this idea.
In addition to context-specific pronounciations I would also add dialect-specific ones. From my experience, some Ukrainain dialects shift stress and vowel quality enough that pairs which might not rhyme in the standard language rhyme more naturally. It's a trick that's used by poets and I wonder how and where do LLMs catch that. (and are they even capable of doing that with less popular languages)