Making medical info accessible as audio

We were asked to take written content that people either didn’t have the time or inclination to read and make it accessible as podcasts.

Converting medical advice for parents of children with a rare disease is challenging. There’s a reason why complexity exists in content: it’s a complicated topic and there are many nuances. We needed to find a way to make it accurate at the same time as being easy to follow.

Using AI to make the audio added three more problems. It couldn’t be AI slop: whatever we produced had to be considered and useful. It couldn’t sound robotic or fall into the uncanny valley, because people would switch off. And it couldn’t mislead anyone by suggesting the voices had lived experience.

Audio has one big advantage: it forces you to simplify. In a printed document it’s easy to leave in extra detail and callout boxes. They might be useful, but they add to the reader’s load. Audio has no room for them, because everything has to be said in order, one thing at a time.

Print lets you get away with long words and complicated sentences. Read aloud, they sound odd, which is one reason some audiobooks are hard going. Speaking forces simpler words and simpler sentences.

The flip side is that simplifying risks losing the richness of the advice. To guard against that, we built a content management system for the audio, with the script pipeline running inside it. The pipeline takes a publication from the charity, pulls out its research foundations and organises them into the angle to take, the evidence that supports it, the anchors that will make sense to people living with the condition, how to make it relevant to listeners, and any direct quotes from the material or from experts that must be included. Once a script is written, a verification loop checks that everything said traces back to the publication and that its key points are covered. AI slop is spreading across the internet, helped along by AI companies overpromising and under delivering. No one, let alone a time-pressed parent, will listen to it.

When we think about radio we think about the voices, and rarely about the structure underneath. Most of us can hear the Today programme in our heads, and we know it feels different from a podcast made by people who are passionate about their subject. One leans towards received pronunciation and careful, correct English. The other is more likely to have regional accents and sound like people talking normally. Both are carefully made and both are worth listening to. So, at the start we didn’t know whether to go formal or informal, and the only way to find out was to try both. We tested six formats, from formal to informal, and around 50 voices to see how well AI could make people sound like people. We worked through it with the charity, round by round.

The uncanny valley is a close relative of slop: something almost human, but not quite. Generated audio can sound too clean. When people talk, even on Radio 4, they talk over each other, make small noises of agreement and finish each other’s sentences. Record yourself chatting with friends and it will sound almost incoherent, because so much is left unsaid or half-said. AI audio tends to the other extreme. Lay individual clips end to end with no overlap, and after a few minutes it becomes uncomfortable, even if you can’t say why. So we mixed in the “mm”s and “yeah”s between the two speakers, let them overlap, and added subtle background noise so they sounded like they were in the same room.

Sounding human is not the same as pretending to be human. The voices aren’t connected to any real person. They don’t, and can’t, have lived experience, and that matters even more when the listeners are parents of children with Duchenne. So the rules around script generation stop the characters from ever saying anything that suggests they’ve been through it themselves. The charity can now take any publication, run it through the pipeline and see every decision made along the way. They can change anything they want before it’s turned into audio.

Podcasts reach people who wouldn’t sit down with a booklet, and fit around everything else. It’s easy to listen while you’re cooking, much harder to read. You can listen to the finished episodes on Spotify here https://open.spotify.com/show/21IvWTQgAfm6E6EUCS1TbR