A growing number of radio companies are experimenting with artificial intelligence, but German broadcaster Antenne Deutschland got out in front earlier this year when it launched a streaming radio station fully programmed with AI and voiced by a synthetic air talent named kAI.
The channel, called Absolut Radio AI, uses technology from Radio.Cloud. Its cloud-native software creates a schedule, programs the breaks and creates voice tracks with a feature called Voicetrack.ai. The streamed station targets listeners ages 14 to 49 years with pop and dance music.
“kAI not only gets you through the day in a good mood,” its website promises. “He also has a lot to say. He gently introduces you to the topic of artificial intelligence and explains to you how versatile an AI can be used, how it works and where its limits are.” You can listen here.
Then in August, Antenne Deutschland put the AI format on a local DAB+ channel in Braunschweig in north-central Germany. It’s believed to be the first over-the-air radio broadcast station in the world programmed and voiced full-time with artificial intelligence.
To find out how all this is going, we spoke with Antenne Deutschland Managing Director Mirko Drenger, Program Director Tina Zacher and Radio.Cloud CEO and founder Christian Brenner.
Radio World: Describe the scope of the organization.
Mirko Drenger: There are two platforms for national DAB+ in Germany, each with 16 stations. Antenne Deutschland runs one of the platforms and six of its stations, branded Absolut Radio. We also have many online audio streams such as Absolut Rock, Absolut 80s and Absolut Radio AI. And we sell ad space for ourselves and for third-party stations.
RW: Tina did you have concerns about working with the AI at first?
Tina Zacher: We thought the same things other people do. How would it do on the air? But we trusted in the technology, in Christian and his team. And we have a lot of people at Antenne Deutschland who are interested in AI. We knew it was a risk but we wanted to take it.
RW: As a programmer, do you worry that AI technology would be seen as a replacement for live talent?
Zacher: I don’t think it will. Human beings will always reach out for other human beings — that’s our thing, we want to be where other people are. But the AI makes work much easier in some ways, especially for radio creators.
There are tasks you love but there are many others you hate to do over and over every day. Now we have AI to do those tasks, leaving us more time to create. But I don’t think AI will replace radio hosts in the next 10 years.
RW: Mirko, what’s the business justification for exploring this?
Drenger: For me, the headline is not that it replaces us. It makes us smarter and faster. It’s not about saving money. You still have to write program-specific prompts, develop your own logic and develop it further. It’s about creating more content, faster and smarter. It’s about fun and creativity.
We’re also helping our employees “skill up.” We encourage them to learn because more and more companies like Microsoft and Hewlett-Packard will require these skills. So we’re not just allowing AI; we really want them to learn it, which becomes a big advantage for us.
We’re just at the beginning of the conversation — we have six radio stations with hosts, yet no one is hosting at night, and we have many online audio streams that are not hosted. I don’t know what the future will bring, but we’re not replacing hosts; it’s an add-on.
Brenner: Among customers I’ve spoken with, no one wants to replace people. They want to add it to nonstop music channels on DAB or web radio channels, or start new ones like Antenne Deutschland did. In the U.S. maybe an HD channel that’s currently just playing music can be turned into a hosted channel.
One customer who is about to launch a 24/7 channel in radio will use it with music and jingles and create chart shows from the ’80s. It will automatically create the shows pulling the chart data from the internet.
We are often surprised how good the AI is; but it is not as good as a creative host. Maybe it’s like a boring host who only tells you the time and the next song. If you want a really creative morning show, we’re still far away, and maybe will be forever. But customers want to get experience with the technology.
RW: How does the system work?
Christian Brenner: Everything that comes before the language model is from Radio.Cloud. We have approximately 4,500 prompts and thousands of news sources. We send a smart query to the language model, ChatGPT 4.0 in Azure OpenAI. We doublecheck results with Google Bard. Some answers will need corrections or optimization, or we’ll drop an answer and do it again.
Then we send it to one of several providers of text to speech. There’s additional audio processing, then we integrate the audio and playout. The process is fully automated.
We recently calculated that we have already created 45,000 breaks for Absolut Radio AI and seven other client channels, most of which aren’t public yet. The technology is working reliably on overnights, on the weekend and with weather reports. It’s really automated.
RW: So no one is typing a script, right? The content is generated by your back end, then voiced synthetically.
Brenner: At the beginning there’s work to set up everything to train it. But once your clocks are built and your presets are done, you could leave it alone, and there is no day-to-day typing as with some AI platforms. It is possible to run the station fully automatically including hook promotions, music ramps, all the mixing, the whole workflow.
You don’t need to write these prompts yourself. Now there’s a simple setting so you can tell it, for instance, “Talk about the next few songs for 45 seconds.”
RW: Tina does your staff have to go over the content, monitoring it to make sure it’s all OK?
Zacher: No. Our colleagues and others listen to the program constantly. But we knew that the AI would make some mistakes. This was okay. It even adds a little bit of authenticity. We don’t have three employees listening eight hours and monitoring every break.
Brenner: We’re creating the breaks 24 hours ahead. It takes about five minutes to scroll through all of them. The staff can log in to see all the breaks for today and tomorrow, and on the weekend they can see three days ahead. They can scroll line by line through the playlist; below each voice break they can read and review the text. The only thing they can’t is the weather report, which is updated 10 minutes before broadcasting.
You’re aware that GPT-4 has problems with hallucinations or inventing things. So if the next song is by Madonna and we tell it “Speak about Madonna using this text from Wikipedia,” we also tell it, “Don’t invent anything. Just take this text and make it sound like a radio host. Speak about this or that, and here are supporting documents; but don’t do your own research.” It makes sense anyway because GPT-4 knows nothing after September of 2021.
Similarly, we take a long-form weather report from the government and GPT makes a nice short version; we don’t ask GPT to create a forecast itself.
Google Bard checks for anything that isn’t appropriate. For example, we did a test in which a weather report was hacked to put out political messages instead; the system detected it immediately and blocked the source.
RW: You’ve been up front with listeners that this is moderated by AI. What kind of reactions have you had?
Zacher: We were surprised because the feedback has been very good. They don’t talk about kAI as a machine, they talk about him like a person, they feel connected to him. He even has his own kind of humor, which they like.
[Sign Up for Radio World’s SmartBrief Newsletter]
RW: Mirko what’s stopping you from using it now as a co-host on a broadcast channel?
Drenger: Nothing. We’re just figuring out how it works and where we want to take it.
Brenner: Doing a co-host is much trickier on the technology side. Having two AI voices interacting is easy, but having a human with a co-host who can respond live brings up latency. Right now, even with the best infrastructure, it would take one and a half to two seconds for the co-host to respond. That’s like a satellite interview in the old days. Many text-to-speech providers are working on this, trying to bring latency down below 0.5 seconds.
RW: Now you’ve put Absolut Radio AI on a DAB+ channel between Hannover and Magdeburg, near Volkswagen’s headquarters.
Drenger: The city of Hannover is known as the place in Germany where the clearest German is spoken. So it’s fun that we start with an AI host there. DAB+ is very successful in Germany and we’re waiting for an opportunity to put it on a nationwide broadcast slot.
Brenner: That feed has different imaging and localized weather forecasts from the start. Local news is planned soon.
RW: What else would you want a fellow radio manager to know?
Brenner: The AI is not taking away the creativity of the program director. The PD tells us what they want. The design is in the hands of the organization. You can even dictate whether he should speak to a younger audience or an older one. Further, if Mirko hypothetically needed to program for an Antenne Spain or Antenne Italy, it could be done with a click of a button. The channel would run the same way, with the same cloned voice, but speaking perfect Spanish or English or Italian.
Drenger: I hear people saying AI is a lot of work, it costs a hell of money and people are frightened of it. No. It saves a lot of money, it’s very easy, it’s very fast, and you can learn a lot. And people aren’t frightened. If you set up a team of seven people doing it while everyone else is left out, of course they’d be frightened; you just have to keep them informed and let them participate.
Zacher: Whether you work at a radio station or in any other area, you’ll be working with AI, so don’t fall behind. Use it as much as you can, and take chances. I’d rather be one of the first people to use it than among the last.
RW: Tina you’ve called kAI “a presenter in its infancy.” When you listen to that voice, do you hear a human? Would you be able to tell that it’s synthetic?
Zacher: I don’t hear a human, though maybe I’m not the right listener because I know it’s AI, and I know Jott Lischka, the person from whom the voice was cloned. But I had a call from a colleague at a music label, and he said, “You still have a person hosting the show, right?” I said, “No, it’s a clone, a synthetic voice,” and he said, “I heard it in the background in my office and I could swear that it was a human being.”
Brenner: We also find that when there is music under the voice, it’s harder to detect than when you hear just the dry voice.
The voice was cloned from Jott Lischka, an employee of Radio.Cloud, who is a former radio host. He called his mum and told her to listen to the stream, and she said, “Oh, you’re back on the radio.” She even told him she heard him make a mistake. Every two or three hours there’s still a mistake — maybe an English artist’s name is pronounced as German. We’re still working on those small errors.
But when even the mom of the host says, “Oh, you’re back on the radio,” that for me is the proof.
Zacher: And as a program director, I can tell you, I have some human employees who don’t speak the name of the of the English singer in the correct way!
Listen to kAI. You can hear the stream of Absolut Radio AI here.
[Read More Radio World Stories About Artificial Intelligence]
