AI is studying to talk African languages, thanks to those startups

It’s a wet Saturday morning in Lagos, and the noise inside the general public hospital partitions on town’s coastal edge drowns out the pelting rain. In one among its echoey wards, voices overlap, and weary nurses attempt to handle paper trails that pile up sooner than they’ll course of them. For many years, this scene has been typical: overworked workers hunched over recordsdata, struggling to juggle affected person care with burdensome documentation.

However one thing is altering. In a single consulting room, a younger physician in a lab coat finishes diagnosing a affected person and as an alternative of scribbling right into a paper chart, he speaks. “Fifty-year-old male, presenting with fever, cough, and fatigue. Suspected tuberculosis. Begin therapy protocol.” Inside seconds, his notes seem, transcribed verbatim on his display, full with punctuation.

This shift is powered by Intron, a startup utilizing voice-based synthetic intelligence (AI) to assist healthcare professionals enter medical information utilizing speech.

The startup, which launched in 2020, was based by Tobi Olatunji, a medical physician with a decade of expertise. In lots of African international locations, docs like Olatunji attend to a whole bunch of sufferers every day and take care of loads of paperwork. Intron Well being cuts the time docs spend writing a affected person’s prognosis. Medical doctors can enter sufferers’ medical information, and generate affected person studies by voice instructions. Crucially, Intron’s AI understands Nigerian accents.

Intron is a part of a wave of African startups constructing synthetic intelligence instruments that talk to the continent’s realities. Past native accents, its voice-to-text system understands medical terminology, permitting overworked docs in Nigeria and Kenya to dictate affected person information hands-free, in real-time. Throughout sectors, homegrown AI options like this are bridging long-standing gaps: serving to healthcare staff clear documentation backlogs, enabling buyer assist in indigenous languages, transcribing native court docket proceedings, and making radio broadcasts extra inclusive. As international fashions wrestle with the underrepresentation of African information, homegrown ventures like Intron are constructing instruments skilled on native voices to shut service gaps in medication, schooling, agriculture, and media.

Whereas AI functions on the continent have been affected by insufficient information, Intron’s speech-to-text AI transcription instrument accounts for a lot of African accents. Olatunji says the datasets are skilled on over 3.5 million audio clips throughout African languages, making lodging for 288 accents. The corporate presently cares for greater than 56,000 sufferers throughout over 30 private and non-private hospitals in Nigeria and Kenya, together with the College School Hospital, Ibadan, Aminu Kano Educating Hospital (AKTH), Kano, Babcock Educating Hospital, Ogun, and Meridian Well being Group, Nairobi.

Intron, which has expanded its AI fashions to energy courtrooms and name centres throughout Africa, claims it has helped scale back the turnaround time for radiology reporting on the College School Hospital, Ibadan, by 99.3% from 48 hours to twenty minutes.

Intron isn’t alone. Enter Spitch AI

Past Intron, a rising crop of startups throughout Africa is constructing foundational voice AI, tuned to the continent’s linguistic realities. One among them is Spitch AI, a developer-first audio AI platform. 

When Temi Babs realised that OpenAI’s Whisper “wasn’t well-tuned to African voices,” the previous engineering pupil deserted his note-taking app and got down to construct the lacking layer himself. “African languages weren’t being catered to by these giant fashions, so we pivoted to voice AI for Africa,” he says, recalling the October 2024 launch of SpitchAI.

The startup had a transparent mission: make AI converse the languages of Africans, in each instructions: speech-to-text and text-to-speech. Slightly than construct end-user apps, the Lagos startup sits on the backside of the stack, providing easy APIs and SDKs that allow any crew plug local-language voice capabilities into name centres, media instruments, or studying platforms, with out the necessity for Machine Studying experience. 

Beginning with Yoruba, the crew has since added Hausa, Igbo, Nigerian-accented English and, after demand from East Africa, Amharic. These rails now energy the whole lot from multilingual customer-support hotlines to Nollywood studios that generate artificial dialogue as an alternative of hiring voice actors. “Producers select the voices they need, now they’ll produce their motion pictures in a shorter time and spend much less cash,” stated Babs. 

Builders purchase pay-as-you-go credit to make use of the speech-to-text and text-to-speech APIs, whereas giant companies pay bespoke charges for tailor-made fashions. “We value on a case-by-case foundation: builders go to our portal to purchase credit and make API calls, enterprises need a customized resolution,” Babs explains.  

The Nigerian authorities is within the race too

Whereas personal corporations like Intron and Spitch AI are constructing synthetic intelligence instruments that talk to the continent’s realities, the Nigerian authorities has quietly entered the race to construct a foundational AI mannequin. 

Final 12 months, the Federal Ministry of Communications, Innovation, and Digital Financial system commissioned Awarri, a Lagos-based startup, to develop what would be the nation’s first government-backed giant language mannequin (LLM). The mannequin, now nearing completion, is a direct try to assist improve the illustration of Nigerian languages within the synthetic intelligence programs being constructed all over the world, in line with Nigeria’s tech minister, Bosun Tijani.

Co-founded by entrepreneurs Silas Adekunle and Eniola Edun in 2019, Awarri is constructing the nation’s first LLM mannequin skilled on 5 low-resource Nigerian languages and accented English, in partnership with Information.org.

In November 2023, the Lagos-headquartered startup launched an information annotation lab poised to be an AI expertise improvement hub. The lab employs over 100 staff, who’re chargeable for gathering and annotating information, creating language fashions, and growing AI apps.

Awarri additionally launched LangEasy in April 2024, a platform that allows anybody with a smartphone to contribute to coaching the mannequin by way of voice and textual content inputs. LangEasy offers customers sentences to learn out loud, and asks them to save lots of the audio on the app. The app will assist crowdsource information for Awarri’s LLM, in line with the startup founder. 

Constructing Nigerian AI is something however straightforward

Constructing Nigerian AI is something however straightforward. Coaching information stays scarce, particularly for indigenous Nigerian languages with low digital footprints. Infrastructure boundaries proceed to gradual progress, and the effectiveness of those AI instruments will depend on how properly they seize Nigeria’s full linguistic and cultural variety.

“Constructing in AI requires two main issues: computing and information. That pipeline doesn’t exist in Nigeria at scale,” stated Babs. To get round it, SpitchAI has stitched collectively an inside data-collection and labelling pipeline and leans on companions for proprietary corpora. “We now have our personal information infrastructure plus some actually good information companions that assist us get the high-quality units we’d like,” Babs provides. 

For Intron, the method includes managing over 3,000 energetic contributors, amassing hours of annotated voice information, paying staff throughout borders, and verifying each utterance for training-grade high quality.

“Individuals suppose it’s nearly fashions, however the actual moat is operations. You’re managing large linguistic variety—Igbo, Hausa, Wolof—and paying hundreds of contributors. That’s not only a tech downside. It’s logistics, high quality management, finance, and belief at scale,” Olatunji notes. 

On the computing facet, prices are brutal. Coaching a single mannequin takes weeks of GPU time and cloud orchestration. Regardless of credit from AWS, GCP, and Azure, these budgets dry up quick. “We’ve burned by way of $100K in AWS credit already. And that’s only one coaching cycle. You want larger GPUs, sooner clusters, each experiment prices actual cash,” stated Olatunji.

However infrastructure is simply half the battle.

In contrast to many AI startups that wrap open-source fashions, Intron builds from scratch. That requires actual analysis expertise, and in Africa, that’s onerous to search out, tougher to coach, and almost inconceivable to retain.

“Zuckerberg is hiring ML researchers at $200M. How do you compete with that? We mentor our crew internally, however even that isn’t scalable anymore,” stated Olatunji.

The crew now runs an open-source analysis group of 30–40 PhDs globally, publishes papers, and makes use of that community to rent strategically, avoiding inflated U.S. salaries whereas nonetheless accessing prime  expertise.

Even after coaching, Intron spent over a 12 months determining the best way to serve fashions to simply 100 customers reliably. “Individuals suppose calling an API is scaling. It’s not. You’re coping with mannequin orchestration, concurrency, GPU latency, price optimization. At AWS, we had 30 engineers simply managing that layer.” Intron needed to compress that right into a 30-person firm.

The chance: a voice-first Africa, outlined by locals

Regardless of the headwinds, the trail is obvious. Intron and Spitch are two of the earliest and most critical contenders within the race to outline what voice AI ought to appear to be for African languages. Whereas international gamers proceed to prioritise scale over specificity, each startups are digging into the onerous, messy, and worthwhile work of constructing AI that really understands the continent.

They usually’re already seeing traction.

Intron is touchdown enterprise offers throughout healthcare, authorized companies, and authorities, with purchasers like Ogun State Judiciary changing legacy voice programs and hospitals adopting its transcription stack for scientific recordkeeping. The corporate can be powering voice authentication for fintechs and dealing on voice-enabled banking and KYC flows, permitting customers to conduct transactions by way of pure speech. 

“We would like a future the place somebody says, ‘Ship ₦5,000 to my brother,’ and it simply works in Yoruba, Swahili, or African French,” says Olatunji. “We’re not simply doing AI. We’re designing usability for the place the infrastructure really is.”

Spitch, in the meantime, is positioning itself as developer infrastructure for voice, providing a no-code conversational agent builder (presently in personal beta) and supporting early customers in healthcare, authorized tech, buyer assist, and leisure.

“We wish to sit on the backside of the stack,” says Babs. “Let different industries innovate on prime of us, whether or not it’s a financial institution in Lagos or a media startup in Nairobi. That’s how we scale.”

This platform-first method offers Spitch extra floor space than vertical-focused startups and aligns with investor expectations in a nascent market. “We’re not chasing fast income,” Babs provides. “The actual upside is in attain. Savvy buyers know it is a long-term sport.”

If each information engines scale as promised, Intron and Spitch may outline the default voice layer for the area lengthy earlier than Massive Tech learns to roll its r’s in Yoruba.

This report was produced with assist from the Centre for Journalism Innovation and Growth (CJID) and Luminate.

Mark your calendars! Moonshot by TechCabal is again in Lagos on October 15–16! Be part of Africa’s prime founders, creatives & tech leaders for two days of keynotes, mixers & future-forward concepts. Early chook tickets now 20% off—don’t snooze! moonshot.techcabal.com