A podcast in 10 languages – the multilingual AI experiment of Louie Media and ELLE

Globally, 87% of young women have felt a negative impact of online misinformation. An experiment by Plan International found that around 53% of social media posts women receive about sexual health are negative. Young women often don’t have access to a reliable online source that answers their physical, mental and sexual health queries. To fill this vacuum, two organisations co-produced their podcast ‘Faut que je te dise’ (“I've Got Something to Tell You”), marking ELLE’s 80th anniversary. 

Supported by Plan International and the Mouvement de la French Touch, this feminist podcast addresses sensitive and intimate questions for girls, helping them make more informed choices. Originating in France, ELLE is a fashion and lifestyle magazine for women globally. However, ELLE France’s venture into podcasting is limited to only six podcasts, covering topics like celebrity news and conversations between mother-daughter duos. Louie Media, on the other hand, is a leading narrative and journalistic podcast studio based in France.

Instead of using AI to translate existing catalogues of podcasts, Louie Media and ELLE decided to create a podcast meant for AI. Charlotte Pudlowski, Louie Media's President and co-founder, says, “We asked ourselves, what could we do with AI that we can't do without? Also, what was really needed right now was to inform young girls from all over the world about their rights, how their bodies work, and to provide it fact-checked in one place.”

The collaboration was made possible due to the organisations’ pre-existing relationship. Louie Media’s creative agency has produced many of ELLE’s podcasts. Louie Media enjoys a younger demographic of female listeners. 82% of their audience base is women, with 70% of them being under 34 years of age. In contrast, ELLE has a demographic of older women. Pudlowski says, “While our podcast is aimed at teenagers, the content is such that parents and mothers especially could pass on to their children.”

Pudlowski explains that the entire process was designed knowing AI tools would be used. The multilingual aspect of it made the team aware that they were addressing an international audience. The script was hence made more general, especially when citing laws. The writing process was done in French with journalist Marine Revol. Then the script was fact-checked and sent to a team of translators. “We had a lot of adjustments to make in terms of what kind of translation we want. Since we're addressing a young audience, how much do we want to pick words that they use” Pudlowski recalls. Once the work on the script was completed, the editorial team passed it over to Fameplay.AI. 

Based in the Czech Republic, Fameplay.AI is a specialist in generative AI for audiovisual & sound production. It was Fameplay.AI that helped turn journalist Revol’s voice into 9 additional languages. With the option of going beyond the original French recording, the team at Louie Media and ELLE had to pick the target languages. Pudlowski shares that the “languages were selected based on a mixture of editorial considerations and technical considerations.” 

“Italy is not a huge podcast market, but we felt it was interesting because it's both a country that's very eager to discover feminist content, but one that also has the far right in the government,” explains Pudlowski. The podcast was translated and voice-cloned into nine additional languages - English, Spanish, Italian, German, Portuguese, Arabic, Polish, Russian and Czech. 

It took the Fameplay.AI team 6 months to produce all 20 episodes in each of the 9 languages. The time was utilised on voice cloning and recording, but also on incorporating feedback from the many proof-listeners they had for each language.

If the team were following the recording in the traditional way, i.e. having a human record it, the total time needed to record the podcast in 9 additional languages would be 172,800 minutes. Fameplay.AI states that using AI, the time was reduced to 2,160 minutes, with each episode requiring 108 minutes of production time instead of the expected 8,640 minutes of recording. This increased the efficiency of recording by 80 times. 

Despite enjoying an oasis of opportunity in turning this podcast into multiple languages, there were many challenges that the team had to face to make this multilingual global podcast come true.

Fameplay.AI achieved voice cloning by utilising a mix of internal and external tools. Ivana Zalesak, the project lead in localisation at Fameplay.AI says, “The AI is well trained in English. It's one of the most spoken languages in the world and the tool itself was US-made.” But apart from English, the tool showed multiple discrepancies in other languages. Zalesak says that Arabic happened to be their most difficult language to master given its 30+ dialects. 

They overcame the challenge by training the tool themselves. “We had to create the accents ourselves and train it”, says Zalesak. She notes that the difficulty was particularly high for languages like Czech and Polish. In contrast, Italian and Portuguese fared better for the Fameplay.AI team. 

Apart from language, dialects and accents also posed a challenge for the team. Pudlowski recalls, “For countries like Italy or Brazil, especially Brazil, given its size, they have very different accents depending on where you are.” This led to the problem of selecting which accent to feature in the podcast as well as the political and regional consequences of how it represents the country.  

The topic of accent was also a problem for the Fameplay.AI team. Zalesak explains that post-recording, the proof-listeners offered different opinions on the podcast. Leading to “the AI tool changing accent through the episodes, which for us [non-native speakers] was very difficult to hear. We would then have to regenerate it so that the AI kept the same accent.”

Despite the amazing results of cloning journalist Revol’s voice, the AI was insufficient at adding warmth to the voice. “We're aware that in some episodes, you sense something not entirely natural. And some of the versions of the podcast are more convincing than others,” says Pudlowski. The AI-generated synthetic voice was particularly lacking compared to French, a language more vocal and dynamic than any other. The Fameplay.AI team notes that cloning the voice remotely affected the voice quality in the podcast.  

Despite the few limitations of the AI tools, the experiment was successful for the creators. “It was very important for us to do it as soon as possible and not wait for a more efficient tool. We wanted to launch something with AI. We had an idea for the podcast, and so we worked on it. We decided to see how it unfolds. And we're actually really happy with the results,” concludes Pudlowski.

This multilingual podcast gained 150,000 listens in its first month. 40% of their listeners came from outside France. Pudlowski adds that the team had not leveraged any marketing in countries apart from France. They say they also faced resistance from certain platforms due to their podcast having "woke content." Their latest numbers indicate 300K listens. Louie Media has declined sharing their detailed audience numbers with The Fix. 

This podcast showcases the vast possibilities AI tools provide for media creators. But also the many challenges and limitations that come with using synthetic voices. Yet, the podcast is notable from a technical perspective for being available in 10 languages. The experiment of Louie Media and ELLE, with the help of Fameplay.AI proved in a tangible way the scope and impact that AI tools can have when used in a meaningful and creative way. 

Instead of replacing AI tools with human effort, the team kept human creativity and judgment at the core of the process. AI was used not as a substitute but as a tool to expand their capability. Pudlowski reflects, “We realised that if we had to find humans for the nine languages to work with us on all the episodes at the same time, we would not have been able to do it at all. It was something we could have done only with AI.” 

Source of the cover photo: Jacob Hodgson via Unsplash