Novel Approach to Live Captioning Through Re-speaking: Tailoring Speech Recognition to Re-speaker's Needs

摘要

A novel approach to the live captioning through re-speaking is introduced in this paper. We describe our concept of re-speaking using only one re-speaker with enhanced re-speaker tasks fully integrated to the recognition system and captioning software. New techniques for instant correction of recognition output, punctuation mark introduction or new word addition are presented. Our real-time recognition system of the Czech language with a vocabulary containing more than one million words is described and an architecture of captioning system that we operate is illustrated. Last part of the paper is dedicated to the re-speaker traininig methodology and a three-level evaluation method of final re-speaker's skills is proposed.