The golden rule about accents is that everyone’s got one. However, although the fodder of thousands of comedic impressions, there’s also a serious side to accents.
Studies in the US, for instance, suggest that some regional accents are slowly fading. Subtle shifts have been observed in places like Texas and Boston, where traditional speech patterns are becoming less distinct. Linguists often attribute this to a process known as levelling, where speakers adapt their accents when they move to new regions to be better understood or to avoid prejudices tied to how they sound.
With the advent of AI language tutors, a new question is emerging as to whether these too will “level” accents to the point where regional or less-spoken accents are in danger. Could language learning apps accelerate the demise of an accent or dialect?
The Risk of Standardization
As millions turn to digital tools to learn new languages, concerns are growing that these platforms may favour “idealized” or standardized accents over regional ones. In doing so they may unintentionally contribute to the erosion of linguistic diversity.
One recent study by SAS examined how regional dialects and phrases in the UK are disappearing as communication becomes increasingly standardized through technology, social media, and AI. Analyzing 100 regional terms using Google Books data from 1919 to 2019, researchers found dramatic declines in the use of words such as the Cornish greeting “ansum,” which dropped by 97%, and the northern English word “scran” (meaning food) which fell by more than 96%.
Researchers argued that AI tools and online communication are contributing to a “digital dialect” that prioritizes standardized speech over local variation, potentially accelerating the erosion of regional language identities.
Some popular language-learning apps are tackling the accent issue by using as much data as possible to understand accents from around the world. Ivan Crewkov, the founder of Buddy AI, a conversational AI tutor for kids with over 76 million downloads, told TechCrunch that by developing the app by using over 25,000 hours of children’s speech in their dataset, it highlighted the importance of accent distinction at the point of being able to understand each student.
“We are trying to understand a 4-year-old Brazilian girl who is trying to say her first words in English at the same time as a 4-year-old Arabic girl from Saudi Arabia,” Crewkov said, adding that the students came from backgrounds with “[c]ompletely different accents and completely different languages.”
If learning apps could recognize accents of students, could this also be taught to preserve accent or dialect diversity? Some founders are considering it and mindfully building their AI tutors to counteract the effects on regional accents. Samuel Bissegger is one such example who’s thinking of the best way to promote accent diversity in language education.
“For us this is always something super important,” Bissegger told Startup Beat, adding that, “We also have quite an educational perspective on how to teach a language, which includes complexity, length of responses, sentence structure, word choice, and accents.”
Bissegger’s sensitivity to accents is shaped by his Swiss background, where linguistic diversity is deeply embedded in daily life. Switzerland has four national languages—Romansh, Italian, French, and German—but within those languages exist a wide array of regional variations.
“There’s even this Swiss German term–Kantönligeist–where basically each canton ribs the other just to mock them a little. This also helps us understand how important it is that even within certain languages there is a huge variety, and we also try whenever we can to make sure that we represent them.”
Yet this richness presents a challenge for language learning apps. Digital platforms often rely on consistency and clarity, which can push developers toward standardized accents that are easier to teach and understand at scale.
The Challenge of Teaching Accents with AI
For companies like Univerbal, the issue isn’t a lack of intent but a question of execution. Representing accents accurately—especially in AI-driven systems—is technically complex. According to Bissegger, quality is the deciding factor.
“It’s super important that we can also provide high quality,” Bissegger said, noting that multiple experiments in his native Swiss German haven’t provided the outcomes they were looking for.
“So far the outcome that is generated doesn’t really work for us because it’s a mixture between multiple different accents, and thereby people get more confused rather than learning something new.”
The problem here is not just about reproducing sounds, as a stand up comedian might do when performing impressions, but capturing the underlying structure of how an accent works. Without this learners may end up with a hybrid that doesn’t reflect any community in the real world.
“You have to understand the mechanics behind it to properly speak the accent. Therefore we always say we first have this kind of quality assessment. If the output doesn’t reach a certain quality level, we’d rather not provide it and wait until certain models get better to then try again,” Bissegger told us.
This cautious approach highlights a broader dilemma in language technology: whether it is better to offer a limited but accurate representation of a language, or a broader but potentially flawed one. For now, many developers are choosing the former, even if it comes at the expense of regional accents.
In this regard, whether future language app developers all speak the same language may depend on striking this balance: between embracing the efficiency of standardization while finding ways to reflect the messy, vibrant diversity of how people actually communicate with each other.
This article is a part of our series on the confluence of startups and ethics. If you have a take on a particularly spicy moral conundrum in the world of startups, drop us a line at [email protected].
