Visualizing punctuation restoration in speech transcripts with prosograph

Abstract

We have developed a neural architecture that tests the effect of lexical, morphosyntactic and prosodic features in restoring punctuation in speech transcriptions. Having outperformed a baseline model in terms of precision and recall, we further extend our performance tests by attaching it in a speech recognition pipeline. The visual and interactive testing environment that we prepared helps us observe how our models generalizes in unseen data and also plan our next steps for improvement.