For many speech recognition systems, that switch creates a problem. The English gets transcribed, while the Swahili disappears. For businesses relying on those transcripts, that can mean losing an important part of the conversation.
Lagos-based voice technology startup Intron is trying to solve that problem with its latest AI models, Sahara v2.5.
The new release adds code-switching support to 12 African languages, including Zulu, Hausa, Swahili and Luganda. Code-switching refers to moving between languages within the same conversation or even the same sentence.
An AI that can handle three languages
Sahara v2.5 also includes what Intron describes as the first African trilingual speech recognition model.
The model is designed for conversations that move between Kinyarwanda, English and French, a common combination in Rwanda’s professional environment.
Kinyarwanda is Rwanda’s national language and is spoken by almost the entire population, while English and French are also widely used.
Intron says it has filed US patents covering the technology behind the models.
The company’s broader argument is that language switching is not a minor problem that can be solved by simply making existing speech models bigger.
A model may perform well in English and Swahili separately but struggle when the two languages appear in the same sentence.
Intron believes the solution requires training models specifically on mixed-language speech, with data and testing designed around how Africans actually communicate.
How well does it work?
On Intron’s internal tests, Sahara v2.5 recorded an average word error rate of 34.3% across 12 languages of code-switched African speech.
That compared with 53.8% for Google’s Gemini 3.6, according to Intron.
The company also says Sahara outperformed Gemini, ElevenLabs and Meta across all 12 languages it tested.
But the numbers also show that the technology is far from perfect. A 34.3% word error rate means the system still gets roughly one out of every three words wrong.
For Intron, however, the improvement shows that models trained specifically for African speech can perform better on local language patterns than general-purpose systems.
Why is switching languages so difficult?
Understanding mixed-language speech is harder than reading mixed-language text.
With written text, spaces and punctuation give an AI clues about where words begin and end. Audio provides none of those signals.
The system has to figure out the words, accent, context and language at the same time. It also has to recognise when one language ends and another begins.
The switch could happen between sentences, in the middle of a sentence, or with just one word.
Many speech systems deal with this by first identifying the language being spoken and then sending each section to a model trained for that particular language.
That can work when someone speaks one language for a long stretch. It becomes much harder when the speaker switches languages for just a word or two.
Intron says Sahara is trained directly on mixed-language speech. This allows it to learn common language combinations and use the context of the entire sentence when deciding what the speaker said.
There is another challenge: tokenisation.
AI models break language into smaller units called tokens. If a model has seen very little African-language data during training, it has fewer tokens for recognising those words.
When it encounters an unfamiliar sound, it may replace it with something that sounds closer to an English or French word—or simply leave it out.
The result can be a word the speaker never said or a missing section of the conversation.
And collecting the right training data is difficult.
Natural conversations where people switch between languages are not easy to collect at scale. Artificially mixing recordings does not always reflect the way people naturally switch languages.
“Code-switching was one of the biggest problems that consistently came up for clients deploying real-world voice AI,” said Tobi Olatunji, Intron’s CEO.
“Africa needs AI built for how Africans really speak. People should not have to translate themselves for a machine, flatten their accent, avoid local expressions or repeat only the English part of what they said.”
Who is using Sahara?
Intron says its technology is now being used for production and research by more than 40 organisations across six countries: Nigeria, Kenya, South Africa, Uganda, Rwanda and Ghana.
One example is Branch International, a fintech lender using Sahara-powered voice agents for loan collections.
According to Intron, the agents recovered more than ₦1.2 million ($891) in delinquent loans in one week. The company also reported stronger performance on loans that had been overdue for more than 356 days, alongside more repayments outside normal working hours.
“Customers engaged naturally even after hours and on weekends,” said Adanne Anene, Head of Product Africa at Branch.
Intron has also expanded Sahara’s text-to-speech capabilities, allowing its voice agents to handle language switching across 13 language pairs.
The company says its models outperformed ElevenLabs and Gemini in nine of those pairs in its internal tests.
The latest release also adds support for Nupe, Kanuri, Nigerian Fulfulde, Tigrinya, Kikuyu, Dholuo and Somali, bringing Sahara’s total language coverage to 31.
For developers building real-time applications, Intron has added streaming speech recognition and streaming text-to-speech to its API. These features can be used for applications such as live captions and voice agents.
From healthcare paperwork to voice AI
Intron was founded in 2020 by Tobi Olatunji, a Nigerian-trained doctor, and Kunle Asekun.
The company initially focused on solving administrative problems in healthcare and raised $1.6 million in pre-seed funding in 2024.
It has since expanded its focus to a much broader problem: making voice AI work across Africa’s many languages, accents and communication styles.
The company has also published a 2026 Africa Voice AI Report, which argues that language data is only one part of the challenge.
Building reliable voice AI also requires research capacity, strong technical infrastructure and people who understand how to deploy these systems in real-world environments.
Healthcare offers a good example.
Ambient medical scribes AI tools that listen to doctor-patient conversations and automatically create clinical notes are already used in parts of the US and Europe.
But deploying the same technology in many African healthcare settings is more complicated.
A single consultation could move between two or three languages.
For Intron, that is precisely the kind of problem voice AI needs to solve if it is going to work for the people who actually use it.