Speech-to-Text for Broadcasters, From Research to Implementation

Speech to text has recently moved from the laboratory to the newsroom as a tool for broadcasters and journalists. Breakthroughs in automatic analysis and improvements in affordability mean that running it at a scale of over hundreds of thousands of hours of content is now feasible. Increases in accuracy mean that users will have a realistic chance of finding what they want in minutes rather than hours, especially in genres such as news or factual content. In this paper, we detail the work the British Broadcasting Corp. (BBC) Research and Development (R & D) have done in this area, and in particular how we built a speech-to-text system using open-source tools, broadcast audio, and subtitles. We detail its accuracy across a range of genres and how it performs on real-life broadcasting problems such as cross-talk, music beds, and laughter. We also explain how an in-house innovation team called BBC News Labs took our work and turned it into digital systems that journalists and program makers can use.