Nigerian voice AI startup Intron AI has launched Sahara v2.5, an upgrade to its voice AI model that adds expanded code-switching capabilities and voice generation for African languages.
The new model can understand conversations where speakers switch between English and African languages, a common feature of everyday conversations across the continent. It also introduces voice generation that can respond in African languages and switch between languages during a conversation.
Intron’s CEO, Tobi Olatunji, says the update was largely driven by feedback from customers using its previous models.
The part of the conversation AI was missing
Intron had already begun exploring code-switching with Sahara v2, but its capabilities were limited. The model supported switching between Swahili and English, leaving out the many other multilingual conversations happening across the continent.
After deploying the technology to customers, Intron began to see the consequences of that limitation. In courtrooms, witnesses could switch from English to a local language while giving testimony. In call centres, customers explaining a frustrating problem might suddenly revert to the language they are most comfortable with. And in hospitals, patients could switch languages when describing symptoms that are difficult to explain. But the Sahara v2, with its limited capabilities, couldn’t keep up.
“The most important part of what they are saying, the AI cannot capture,” Olatunji notes.
That feedback pushed the company to expand what it had started. Sahara v2.5 supports code-switching across about 20 African languages, allowing the model to follow conversations as speakers move between languages rather than treating those moments as errors or gaps.
The company says its focus with this release was particularly on maintaining accuracy at those switching points, where speech recognition models can struggle. The startup plans to release performance benchmarks for the model separately.
“So the 12 languages, (including Zulu, Hausa, Swahili, and Luganda), that are covered in this release span over 10 countries in Africa, where they’ll now be able to use this kind of language.”
AI can now speak back too
Another major addition to Sahara v2.5 is voice generation. While voice AI has become increasingly good at sounding natural in English, generating natural speech in many African languages has remained difficult. Intron says its new model can generate speech in languages including Igbo and Hausa, and enable the generated voice to switch between languages during a conversation.
That could make a difference for businesses trying to build more natural voice experiences. A call centre, for example, could use a voice bot to speak with customers in their preferred language without forcing the entire conversation into English. The same technology could be deployed for public information campaigns, government services, media, or any service where speaking is easier than navigating a screen.
“The model now also supports code-switching,” Olatunji says. “It can actually talk back to you in mixed English and Yoruba, mixed English and Swahili.”
One of the first deployments of the new technology is at Meridian Hospital, Enugu, where conversations between doctors and patients are largely in Igbo.
The hospital had previously used Intron’s English model for dictation. A doctor could finish seeing a patient and dictate their notes into the system. But recording and understanding the conversation itself could remove that extra step. Instead of writing notes from scratch after every consultation, a doctor can review notes generated from the conversation and make minor edits where necessary.
“If I’m seeing 60 patients today, instead of writing notes for 60 patients, I can just edit,” Olatunji says. “It’s a real unlock for doctors.”
Intron believes the same ability to understand natural, mixed-language speech could also make voice AI useful in other sectors, such as the creative industry and agriculture. Olatunji says farmers are a major use case because there’s a significant concentration of AI investment aimed at helping farmers get more from their crops and livestock, and even control pests.
Scaling without trying to sell to everyone
Intron’s expansion across different industries raises an obvious question: how does a startup sell to hospitals, governments, banks, call centres, and others without spreading itself too thin?
Olatunji explains that Intron AI was solely focused on its healthcare offerings, but then realised that its product was serving other verticals like call centres, legal, and government agencies because it understands African names and accents.
“We started by changing our research strategy to ensure that the fundamental model is a good model across board, knowing that different people are going to use this model.”
Intron is increasingly adopting an API-first approach, enabling other companies to integrate its speech and voice capabilities into their own products. The idea is that a company building a financial service, for example, can connect to Intron’s technology rather than Intron having to build and sell a separate product for that industry.
“Our only product is still the underlying technology,” Olatunji maintains. “If you give us voice, we want to give you text. If you give us text, we want to give you voice.”
The company’s direct sales team, meanwhile, focuses on sectors where it sees clearer returns and higher margins, including government partnerships (donor-funded) and financial services.
Voice banking is one area Intron is exploring. Instead of opening an app and navigating menus to complete a transaction, a user could simply say what they want to do. For Olatunji, the value proposition is not necessarily about convincing businesses to adopt AI for its own sake.
“It’s not really about the AI,” he says. “They have value that they want to unlock, and you are just bringing the right technology in.”
The company says it expects to reach profitability next year, driven by its current partnership pipeline, although Olatunji declined to share revenue figures.
Building the technology behind the voice
Getting African voice AI to work at scale requires more than collecting recordings. Intron uses a combination of paid human contributors and a proprietary synthetic data generation system.
Human data helps researchers understand the problem and train models, while synthetic data enables Intron to create more targeted data without relying entirely on the slower, more expensive process of collecting human recordings.
Olatunji says the company has filed a US patent application related to its data-generation technology. But even after a model has been trained, there’s still the challenge of ensuring it delivers the same quality when customers are actually using it.
“A lot of people think it’s just data,” he notes. “There are algorithms, synthetic data, and training recipes.”
The challenge becomes even more complicated as usage grows. High-quality, low-latency voice AI requires significant computing power, and access to high-end GPUs remains both expensive and limited.
Intron is also dealing with the slower side of selling to large organisations. A bank or government agency may be interested in the technology, but turning that interest into a signed deal can take months or even years.
Olatunji says Intron offers a mix of local pricing in countries like Nigeria, Kenya, and South Africa, but it defaults to USD pricing for other markets.
Regarding privacy concerns, he explains that Intron is running a project with a hospital in Puerto Rico in which the AI model is offline. He adds that many countries that want data stored locally don’t have the compute power to back up that regulation.
“In cases where you send to the cloud, you have the option with us and many other providers to ask for zero retention. You keep your data and the results you get, and we have to delete it from our site; it also limits our exposure to legal risks.”
The company is betting that as voice becomes a bigger part of how people interact with technology, models will need to understand the way people actually speak rather than asking people to adapt to how machines understand language.
Olatunji believes voice could eventually become an important interface for people who have been left out of digital services because those services were built around screens, apps, and text, adding that Intron is ensuring that its research, engineering, products, and sales are positioned to capture value.
“If they’re able to do things themselves directly without having to navigate an app or download anything, the economy will start unlocking a huge part of the economy just because we changed the interface,” he notes.
