Most Africans speak at least two languages and often blend them in a single sentence, such as “My loan imekataliwa, but I paid yesterday.” In many speech recognition systems, English is transcribed while Swahili is dropped, leaving out part of the meaning for any company relying on the transcript.

Intron, an Africa-focused voice technology company founded in Lagos, has released Sahara v2.5, a set of models built for that use case. The release adds code-switching support — the ability to follow a speaker moving between languages within a single sentence — across 12 African languages, including Zulu, Hausa, Swahili and Luganda.

The release also includes what Intron describes as the world’s first African trilingual speech recognition model, designed to handle conversations that move between three languages. The model is built for Rwanda, where Kinyarwanda, the national language spoken by almost the entire population, coexists with English and French in everyday professional life. The company says it has filed US patents on the algorithms behind the system.

Intron argues that code-switching is a distinct technical challenge that global labs have often treated as a minor issue. A model can perform well in English, be competent in Swahili, and still fail when the two appear in the same sentence. The company says solving the problem requires dedicated data, dedicated training and dedicated evaluation, and that a smaller company focused on a narrow problem can outperform larger players using global averages.

On Intron’s own benchmarks, Sahara v2.5 recorded an average word error rate of 34.3% across 12 languages of code-switched African speech, compared with 53.8% for Gemini 3.6, a Google AI model. In practical terms, that means Sahara gets about one word wrong in three, while Gemini gets more than one in two.

The company says Sahara outperformed Gemini, ElevenLabs and Meta across all 12 languages it tested. Even so, the system still misidentifies roughly one word in three, underscoring how far speech tools still have to go before they can reliably serve mixed-language conversations that are common in customer support, finance, healthcare and other everyday services.

Founded in 2020 by Tobi Olatunji, a Nigerian-trained doctor, and Kunle Asekun, Intron raised $1.6 million in pre-seed funding in July 2024, led by Microtraction, to address paperwork problems in hospitals. It is now working on voice AI products for use across the continent.

Why code-switching is difficult

Speech is more difficult than text in this context because a written sentence clearly shows where each word begins and ends. An audio model receives only sound. It must identify the sounds, words, accent, context and language simultaneously, with no visible boundary showing where Yoruba ends and English begins. The switch may happen between sentences, inside a sentence or around a single word.

Many systems handle this by first identifying the language and then routing each segment to a monolingual recogniser, a model trained to transcribe one language. That method can work with long, clean segments, but it often breaks down when the switch lasts only one or two words. Intron says Sahara is trained directly on mixed-language speech, allowing the model to treat switching as normal, learn which transitions are possible in each language pair and resolve them using acoustic evidence and context across the full utterance.

The second challenge is tokenisation. Before a model can process speech, it breaks sound into small units called tokens, each matched to a piece of language it has seen before. If African-language forms account for only a small share of the training data, the model has too few African tokens to match against and may force an unfamiliar African sound into the closest English or French token it knows. That can produce a hallucination — a word the speaker never said — or cause part of the speech to disappear from the transcript.

Training data is also limited. Recordings of natural code-switched speech are difficult to collect, and artificially mixed audio does not reflect how people actually switch languages.

“Code-switching was one of the biggest problems that consistently came up for clients deploying real-world voice AI,” Tobi Olatunji, Intron’s CEO, told TechCabal in an interview. “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 model now supports production and research deployments at more than 40 organisations across six countries: Nigeria, Kenya, South Africa, Uganda, Rwanda and Ghana.

One commercial example the company shared is Branch International, a fintech lender. In that deployment, Sahara-powered collections agents recovered more than ₦1.2 million ($891) in delinquent loans in one week, while also recording after-hours and weekend repayments at levels the company described as a record. Intron said the system performed better than human agents on loans more than 356 days overdue.

“Customers engaged naturally even after hours and on weekends,” said Adanne Anene, Head of Product Africa at Branch, in a statement shared with TechCabal.

Intron also said its text-to-speech products, which convert written words into spoken audio, and its voice agents now handle language mixing across 13 language pairs, outperforming ElevenLabs and Gemini in nine of those pairs on the company’s tests. The startup added that its speech recognition now supports Nupe, Kanuri, Nigerian Fulfulde, Tigrinya, Kikuyu, Dholuo and Somali, bringing total language coverage to 31.

The company has also added streaming speech recognition and streaming text-to-speech to its application programming interface (API), giving developers tools for live captions and real-time applications.

Intron has published a 2026 Africa Voice AI Report, arguing that collecting African language data is only one barrier to reliable voice AI. The report says research capacity, the ability to connect systems together and deployment expertise are equally important.

It points to the ambient medical scribe — software that listens to a doctor-patient consultation and automatically drafts the notes — as an example. The technology is widely used in the US and Europe, but is largely impractical in African clinics, where a single consultation may involve two or three languages.

True scale, the company says, requires moving beyond surface-level integrations to robust execution.