A fully AI-run radio? Content on Demand proposes a different, immediately usable approach that builds on what already exists

Not so fast

Nine months have passed: we think we have understood a little more and can correctly position AI as an additional resource, rather than a total change in the approach of broadcasters with consequent staff reductions.

Of course, this is a personal and provisional conclusion because, inevitably, what we thought was feasible at the end of 2022 if not today will likely be possible in the future.

But let's take it step by step.

Patrizia

Let's start with an early adopter, the publishing company behind 70-80.it.  In December 2022 they devise a radio format, the Diario di Patrizia, where the vicissitudes of a 70s/80s teenager are invented and written by GPT 3.5-rurbo through the PeperoniAI front-end.

To get the desired result, they started from a series of stories written by a human copywriter: stories of teenagers, love affairs and family problems.

These first scripts were  writtten by humans and not intended to be broadcast, but to train the AI.

Subsequently, the real scripts were generated by AI: specific instructions were provided to Peperoni, such as:

_Number the stories, identifying them with a progressive number
_Use 80s teen slang
_Write short stories (about ten lines) suitable for radio reading and not too complicated.
_If possible, think of funny stories, even if the goal is not to make people laugh.

The program was a big hit and can be still heard on the station.

 Kai

In Germany someone wanted to go a step further, creating a radio totally managed by an AI, called KAI.

This is the case of Absolute Radio AI, whose manager recently told Newslinet that "Much of the work at Absolut Radio AI is handled by our site, which sends the request to ChatGPT and then re-checks and optimizes the response again.
The platform runs the entire workflow completely on its own;  however, for the first few months humans will check all texts and listen to vocals before broadcast."

Reality Check

We of course performed a reality check by listening to the station. We remember little of the German learned in high school, yet enough to be able to say that the chosen format is quite smart: music with segments where the talk is ... about the music.
In practice, the AI tells something about the tracks or artists on air, which is feasible without too much risk of  hallucinations,  if the model used has been (as we assume) fine-tuned on music history.

The synthetic German voices are credible and  very enjoyable. In fact: surprisingly good.

Content on Demand

This brings us to Content on Demand by 22HBG.

Let's take the  experience, by Radio Nizza, an italian station created to and aimed at the Riviera (Cote d'Azur). In the first two months the station had access to was able to experiment with the new "content on demand" service*.

Their conclusion: Content on Demand can effectively complement the work of editorial staff and presenters for services related to current affairs and possibly programmatically connected to specific feeds.

Content on Demand, let's clarify, is a cloud service created by 22HBG based on what has already been developed for Peperoni and dedicated to creating audio content based on prompts.

As you can see in the screenshot, users can upload a preroll (in our case from the official voice of Radio Nizza) and a backgrounds sound.

Then in the "Context" area users can  insert a text prompt (complex topic that we will return to) to get a complete audio clip created in a few seconds, ready to be verified and aired.

 

An example

Let's explain the process with an example. The editorial staff prepares content every week, the agenda of events on the Côte d'Azur or the news of the day. But it certainly cannot be present 24 hours a day to monitor alerts or specific events, which are however available online. Such as in this tweet by the public transport of Nizza, "Ligne d'Azur", warning users of a change in the line 90.

 

Here's the point. By providing Content on Demand with the message (captured via a Python script) and preceding it with a specific prompt (again crafted via Python) we can automatically create a radio segment, sending Content on demand the following complete prompt:

Prompt: Create an urgent message for Radio Nizza informing - translating into Italian - about the following: "En raison d'un éboulement, la ligne 90 est modifiée le jeudi 3 août 2023."

And here is the result aired on August 3 (the content starts at second 14):

The text is: "Good morning to all our listeners. We have an urgent announcement to make regarding line 90. Due to a landslide, line 90 will undergo changes on Thursday, August 3, 2023. We invite you to take this into account for your travel plans. We will update you with further information as soon as it becomes available. Thank you for your attention and have a nice day everyone!"

French version

The same system can automatically create a French version.  No editorial work is  required: it is  enough to specify "in French" in the prompt:

Prompt: Create an urgent message for Radio Nizza informing - in French - the following: "En raison d'un éboulement, la ligne 90 est modifiée le jeudi 3 août 2023."

This is the excellent result

t,

The text:

Chers auditeurs de Radio Nizza, Nous avons une nouvelle urgente à vous annoncer. En raison d'un éboulement, nous devons vous informer que la ligne 90 sera modifiée ce jeudi 3 août 2023. Cette situation imprévue nous oblige à apporter des changements à l'itinéraire habituel de la ligne 90. Des mesures de sécurité sont en cours pour évaluer les dégâts et rétablir la situation normale dans les plus brefs délais. Nous vous recommandons de prendre vos dispositions pour éviter tout désagrément lors de vos déplacements. Il est possible que des retards ou des modifications temporaires soient nécessaires sur cette ligne. Veuillez suivre les instructions données par le personnel de la compagnie de transports en commun pour connaître les alternatives possibles. Nous comprenons que cela peut causer des désagréments et nous nous excusons pour les inconvénients que cela pourrait vous causer. Nous mettons tout en œuvre pour minimiser les perturbations et rétablir la situation normale dans les plus brefs délais. Restez à l'écoute de Radio Nizza pour les dernières mises à jour concernant cette situation. Nous vous tiendrons informés de tout développement ou changement concernant la ligne 90. Nous vous remercions de votre compréhension et de votre coopération. Prenez soin de vous et soyez prudents lors de vos déplacements. C'était un message urgent de Radio Nizza.

Gallicisms

As you can see the resulting message is very long, forcing to shorten it by reiterating the request to Peperoni preceded by "repeat the previous text shortening it by 50%".

The reason is both ridiculous and interesting: the French - we can now affirm from direct experience - tend to be very ceremonious in writing and speaking, turning even short communications into very long speeches. The AI, trained on this style, has therefore excessively lengthened the message, forcing the station to run a second iteration.

English Version

Finally, the same operation with English:

Ongoing Developments

As for Italian voices, the weakest of the three,  we contacted  the head of development at 22HBG who is working on the next version of Content on Demand.

Among other things,  he said, he is implementing the options to change the "mood" or "tone" of the voice: through presets it will therefore be possible to make the artificial voice speak in a neutral way (BBC World Service style, to be clear) or emphatically (as often adopted by presenters in music programs).

Last Note

Last note: did you read above the manager of Absolute Radio AI stating that "for the first few months humans will check all texts"?

Well our advice is to keep checking the output not "for the first few months" but until "future AI systems that emulate the human "super Ego" - additional downstream models that control the AI's work - are able to reject inappropriate, hallucinated or simply "not radio-friendly" content"

(M.H.B. for FM-World)