Universal-2-TF introduces a two-stage neural text formatting model for ASR that combines token classification and sequence-to-sequence approaches to efficiently handle punctuation, capitalization, and text normalization while achieving superior accuracy across diverse domains.
