Syllables as Linguistic Units?

While there continues to be a debate in linguistics and speech processing as to the nature of an atomic unit, in computational approaches the atomic unit is universally taken to be the orthographic, space-demarcated “word”. We argue that for many richly inflected languages such as Indo-European languages, syllablebased approaches together with semantic grounding may provide certain advantages. As a demonstration, we consider a language-acquisition system for Hindi, and propose a text syllabification technique, and show that syllable-based models perform somewhat better than the traditional word-based approach in building up a noun lexicon based on unannotated commentaries on video. We suggest further exploration of this potentially important idea in other domains.