Neural Text Normalization in Speech-to-Text Systems with Rich Features

This paper presents the task of normalizing Vietnamese transcribed texts in Speech-to-Text (STT) systems. The main purpose is to develop a text normalizer that automatically converts proper nouns and other context-specific formatting of the transcription such as dates, time, and numbers into their appropriate expressions. To this end, we propose a solution that exploits deep neural networks with rich features followed by manually designed rules to recognize and then convert these text sequences. We also introduce a new corpus of 13 K spoken sentences to facilitate the process of the text normalization. The experimental results on this corpus are quite promising. The proposed method yields 90.67% in the F1 score in recognizing sequences of texts that need converting. We hope that this initial work will inspire other follow-up research on this important but unexplored problem.