Sanoma uses Azure AI Speech Service for low-cost, automated weather forecasts

The weather in Finland can change without warning; a balmy morning with blue skies might give way to sleet, snow, and harsh winds by lunchtime—which is why Finnish people tend to check the forecast often.  

And because temperatures can vary greatly between the southern and northern regions of the nation, local forecasts are essential. 

As one of the country’s largest media companies, Sanoma reaches millions each day with TV, radio, magazine, and newspaper content. It is adept at serving local audiences, yet its radio stations have until recently only offered national weather forecasts during its news broadcasts.

Sami Virtanen, Sanoma’s Content Insight Strategist for radio, says that while the company was keen to offer more locally focused forecasts, the business case for doing so was not strong enough. 

“It would have been quite expensive, having all these separate weather departments,” he says.   

But things changed radically last year when its two radio stations began airing local forecasts across 26 different regions, three times a day—all using AI technology from Microsoft. “We are now producing 78 additional forecasts a day, quickly and efficiently, for an audience of up to 200,000 people,” Virtanen says.

Each forecast is read by a lifelike synthetic voice created using the text-to-speech feature Custom Neural Voice (CNV), part of Azure AI Speech Service. Sanoma also uses Azure OpenAI’s GPT-4o to prepare the reports with data from the Finnish Meteorological Institute. 

“There was such a large volume of regional weather data that it would be a full-time job to manage it all manually,” says Renny Järvinen, Head of Sound Production at Nelonen Media, a subsidiary of Sanoma. “With AI we were able to fully automate the whole process.”

This transformation means that “Sanoma can offer a better product to listeners at a fraction of what it would have cost to do manually,” Järvinen says. It also creates opportunities to grow advertising revenues; local businesses can pay to be featured on one of the forecasts. And since the initiative is easy to expand, it is an enabler for efficient growth in the future. 

“It’s very scalable, so if we want to make our stations even more local, we have the capability,” Järvinen says. Sanoma also has plans to use CNV for certain news broadcasts.

Building AI initiatives on a foundation of trust 

Sanoma has a long history of keeping the Finnish public informed, with deep institutional expertise in providing accurate information—and multiple brands that have earned the trust of a wide audience. 

Amid concerns about the impact of AI on the media landscape, this is a major asset, says Pauli Tölli, Sanoma’s Chief AI Officer.

“Whatever we are doing with AI, it's good to remember that we started publishing newspapers in 1889,” he says. “We have lots of historical content, which is one reason why our first priority with AI is safe usage and intellectual property rights.”

As well as the weather forecasts, Sanoma has synthetic voices to read its news articles and AI-generated summaries of major events such as the last US election. It also uses AI to suggest headlines that get more clicks, and to save time in the news gathering process by summarizing information from its own published materials. 

Whatever the use case, Sanoma applies the same standards of rigor, transparency, and consistency to AI-generated content as it does to any of its output, he adds. For the weather broadcasts this means, among other things, an announcement informing audiences that they are AI generated. 

“We see that readers and consumers trust our news brands. We need to earn that trust every day,” Tölli says. 

Sanoma conducted a survey before launching the synthetic voice to determine listeners’ attitudes to the change. It found that only 11% of respondents had listened to news or audio articles read by an AI voice. On top of that, three quarters said they associated AI-generated media with artificiality and fakeness. 

Yet when respondents actually heard the synthetic audio samples played by Sanoma, the vast majority had a positive perception. “When they heard the voices, it sounded much better than they had probably expected,” Virtanen says. 

In a follow-up survey this year, Sanoma found audience perceptions of the synthetic broadcasts remained strong. Their Microsoft custom Supi voice outperformed other voices on the market in terms of perceived suitability: 70% said it suited news articles either fairly well or very well. On top of that, people were more likely to think it sounded like a human than a machine when asked to guess based on a sample.

Audiences are growing more familiar with AI in media

The earlier survey also gathered data about people’s perceptions of AI in the media generally. The majority, 57%, had a neutral attitude towards AI-generated media content, while 27% had a negative one. Respondents trusted AI-generated content from established media more than from less established sources online, but the results showed there was still a way to go in earning the trust of the majority.  

By using AI responsibly and transparently, Sanoma is working to build trust over time. “Since we took that survey last year, I think our audience has gotten more used to content generated with AI, and they can see that we have not generated any fake news or taken things in the wrong direction,” Virtanen says. “At the same time, it’s helping to make our products better and improve the audience’s experience.”

Issues are easy to resolve and output is consistent

To develop the AI weather service, Sanoma had to overcome several challenges. One issue was that every forecast had to be the same length so that they would fit into the live national news broadcast. 

“That was critical for our success because they needed to be on air at the same time and they needed to end at the same time,” Järvinen says. “But AI isn’t very good at estimating text length.” 

They solved this issue by designing each local weather report to consist of only five sentences, with each sentence defined by a strict prompt structure.

To train the voice itself, they used a base script from Microsoft and then added some of their own custom phrases—known as “training utterances”—to make it more relevant to weather reporting, amounting to a total of around 750 utterances. 

Another thing that the team found from its survey was that when the voice was based on a well-known broadcasting personality, listeners found it less convincing. “If we use the voice of someone famous, then listeners are always comparing it very closely to their own idea of the real thing,” Virtanen says. 

So instead, they trained the model using voices from previous presenters who were no longer well known.   

“Choosing the right human voice for the base recording was crucial,” Järvinen says. “This is where we have the advantage, because we work with voice and audio for a living. We have a lot of experience with this, and we have access to a wide range of voice talents in the Finnish market.”

During testing, the team noticed some pronunciation issues, which Järvinen says may have been due to the fact that Finnish is not widely spoken and there is consequently less testing with Finnish voices. 

“Microsoft helped us to address this issue promptly, and we received really great support from them.” 

Another benefit of opting for Microsoft, he adds, is that their CNV technology offers a higher level of reliability. 

“With other voices you can’t fully trust the output and you need constant human oversight, but with Microsoft’s neural voice, you don’t have to do that. If there is a problem, you can reproduce it and therefore you can fix it. That’s not the case for competitors, and it’s why this voice is best for us,” he says. “That means once it is fine-tuned, we can rely on it to perform consistently and accurately. So I think the neural voice is essential in a project like this, because the final result must always be high quality and predictable. We can't broadcast an unpredictable voice on radio or TV.”