Ensemble pour le DAB+ urges French government to strengthen public support
LONDON — For much of its existence as an innovative medium, radio has focused primarily on advances in audio technology, including sound recording and editing, transmission and hardware such as microphones. However, as times and the needs of broadcasters have moved on, the scope of research and development has broadened to encompass other associated technologies, such as data. The BBC has been particularly active in this area and this year launched two sound-related pilot projects that are not audio in the accepted sense — one to provide subtitles and transcripts on BBC Sounds programs and podcasts, and the other offering a discovery tool to enable listeners to find new music.
BBC Sounds, the U.K. public broadcaster’s app for all its radio programs, podcasts and related audio, produces approximately 27,000 hours of material a month. Accessing some or all of this can be problematic for those among the 18 million people in the U.K. with some form of hearing impairment.
The BBC has been investigating how to add subtitles to programs without making the process time-consuming or prohibitively expensive.
BBC R&D principal engineer Andrew McParland explains that work had already been going on to create a speech-to-text system, primarily for archive purposes. “What we noticed in recent years, particularly with the release of Whisper AI, is there has been a step change in the performance of these systems,” he says. “This means the quality of the transcripts they produce is much higher, and the word error rate is much lower. That offered an opportunity to automatically generate these transcripts and, potentially, subtitles.”
Automatic speech recognition
Whisper AI was launched in 2022 by OpenAI, the developer of ChatGPT. It is classified as an automatic speech recognition (ASR) neural net system and was trained on 680,000 hours of multilingual and multitask data. As an open-source technology, the BBC sees it as a way to produce an efficient transcription service that would be more accessible to its listeners.
“Putting transcripts on podcasts has been around for a while as an accessibility feature on several different platforms, but what distinguishes BBC Sounds from those other outlets is the massive volume of content we make,” comments Nigel Megitt, BBC executive product manager for access services technology. “The scalability of providing transcripts for that is a big technical challenge. We already had the infrastructure to distribute transcripts, but we needed to do some work to allow our users to view them. The trial project came out of that to see whether the quality was good enough and what challenges were involved.”
The technology has been tested on various programs and podcasts, including BBC Radio 4’s flagship accessibility programs, In Touch, Access All, Profile, Sporting Witness and Economics with Subtitles. The web version of BBC Sounds offers subtitles, while the Android and iOS apps offer full transcripts. The material is generated automatically from speech but then checked by a journalist to ensure it is accurate and makes sense.
The trial began in August 2024 and was scheduled to run for three months. At the time of the interview in October, McParland and Megitt said it was too early to discuss specific points raised by the tests. McParland added that there had been some user feedback, but the development team was still learning. “Even if the trial in its current form shows that the system isn’t quite ready for prime time, what can we learn from it, and what would it take to make it perform better, or do we wait for the technology to improve?”
BBC Orbit music discovery pilot
Radio broadcasters play an important role in finding the next generation of musical talent. Since 2007, the BBC’s Introducing platform has not only discovered up-and-coming British artists but also, most importantly, presented them to listeners. The problem is that the local DJs at the different stations within the four U.K. nations — England, Scotland, Wales and Northern Ireland — have limited time in their programs to showcase emerging singers and bands.
This has led to the testing of BBC Orbit, a prototype “discovery app” that offers samples of tracks by undiscovered artists on a “starfield.” This is a circle on the screen with 10 points, each representing a song that can be heard by dragging a central “joystick” over it. These are not identified by performer name, song title or genre; people listen to snatches of each track until they find something they like. When they do, the song details can be revealed, and people can hear the full track on their preferred music streaming service.
Orbit grew out of a project that began a year ago to develop new audio products for young people. “It was one of the more successful prototypes we tried,” explains Mathieu Triay, principal software engineer with BBC R&D. “What we identified was that discovery wasn’t a problem in itself. People could find content because it was suggested on the streaming platform they used, but if they wanted to find something themselves, that was harder. It seemed like the power dynamic of searching had shifted, and we thought maybe we could do something to shift it back.”
This involved removing some of the biases associated with relying on music genres. “People might say they never listen to hip-hop or electronic music, but once they find one track they like, there’s an entry point allowing them to discover [more],” Triay said.
“Notably, we found that separating genres made less sense with young people because there’s now more overlap between genres.”
The original prototype did not include information about the artists. It offered only what Triay describes as a “visualization,” with a joystick that enabled “blind navigation” until someone found what they liked. This was tested on a focus group of young people who reacted positively, leading to a pilot scheme that has been running since September.
“It seemed like the power dynamic of searching had shifted, and we thought maybe we could do something to shift it back.”
Mathieu Triay, principal software engineer with BBC R&D
Rethinking genres
Orbit is a joint project between BBC R&D and BBC Introducing, whose local presenters provide tracks from new artists in their respective areas. “The DJs are already doing the work, and that’s one of the reasons why Orbit was selected as a project,” comments Rob Scott, development producer at BBC R&D, who liaised with BBC Introducing presenters. “We already have all these humans listening to music, hundreds of tracks a week, and championing new British music. So, we had a system for them [the DJs] to receive music from local artists, listen to it, and then decide whether to put it in their show.”
In bringing together all this material and trying to present it in an easy-to-navigate way, Triay and his colleague, Senior User Experience Designer Andrew Wood, decided that they did need to use genres but in a different way than just labeling something “rock” or “hip-hop.”
“Genres allow us to organize stuff and place things in relation to one another,” says Triay. “We discovered we need to break it down into parts. So, what do you understand when you say ‘rock’? It’s got electric guitars, strong drums, a solo and strong vocals. If we can quantify all of this for every song, then we’d be able to essentially have a set of numerical parameters that we could use to plot things.”
This was done using an open-source C++ library, Essentia, designed for audio analysis and audio-based music information retrieval. It uses a range of machine learning models trained on the various elements of different music types. “It can give us a rating from minus one to one on various criteria, including how aggressive the music is, how instrumental it is, how happy-sounding it is or how sad it is. And we tried over 30 different criteria,” Triay says.
The Orbit project was due to finish by the end of December 2024. Where it can go from there is still being considered, although Scott says it would not become a product or service in its own right because the BBC already has six core digital products, including BBC Sounds. “We will present this mechanism of blind intentional spatial discovery, which Sounds could pick up to focus more exclusively on new music,” he says. “Someone also suggested that Orbit could be a way to find music to use in TV shows. It won’t become its own BBC digital service, but it could go in several directions.”
The author trained as a radio journalist and worked for British Forces Broadcasting Services Radio as a technical operator, producer and presenter before moving into magazine writing during the late 1980s. He recently returned to radio through his involvement in an online station where he lives on the south coast of England.
More stories about how AI is changing radio
The risks of AI in the radio industry
