active

Streaming Transcription

A prototype for transcribing live feeds in real-time, as they happen

Hypothesis

How can we use transcription technology to help journalists cover breaking news stories?

Our Streaming Transcription prototype is a web application that transcribes incoming audio and video feeds in real time.

It's built using a transcription engine trained on BBC data, which has been specifically optimised for handling streaming media.

Screenshot of Streaming Transcription prototype.

The prototype allows users to jump to certain points in the media stream by clicking on words in the transcript

Why?

We already know that there's huge demand in the newsroom for being able to monitor and navigate through audio and video files using automatically generated transcripts. Our work on file-based transcription — where journalists upload a recording and wait for a transcript to be generated — has recently been transferred into production.

We wanted to explore whether we could build a similar tool, but for streaming media feeds.

One of the BBC's top priorities for digital news is being the best at delivering breaking stories. We know that the teams producing our live pages currently need to allocate journalists to manually transcribe quotes from events such as press conferences. We thought: could we make monitoring live event feeds easier?

Luckily, our colleagues in BBC Research and Development have already developed a live transcription engine that's optimised for handling streaming media. However, it hasn't been tested in the newsroom yet.

This project gives us the opportunity to evaluate the speed and accuracy of this streamlined speech-to-text service, and feed back to our colleagues in R&D with our findings.

Next steps

  • We will run a trial with journalists in the newsroom to test whether the speed and quality of the transcription is good enough to be useful when covering breaking news
  • If the speed and accuracy is good enough, we will look to improve the design of our prototype and add additional features before running additional user testing

A special thanks to Matt Haynes, Misa Ogura and Ben Clark from BBC R&D for their work on streaming Kaldi.

Results

  • We've completed a prototype tool which we will test with journalists to understand whether the transcription is fast and accurate enough to be used when covering breaking news.

Careers

Love data and code?

We're looking for talented developers to join our team.