Tag
This paper introduces UGTPhon, a benchmark for grapheme-to-phoneme conversion in user-generated text, and presents a compositional approach that improves performance by leveraging canonical forms.