In 2022, a study conducted by Intron, a Nigerian AI startup, involving Nigerian doctors, found that their average typing speed was just 22 words per minute.
The figure might seem like an unimportant detail, but it points to a much bigger problem in Nigeria’s healthcare system.
Every day, a doctor must listen to a patient describe their symptoms, ask follow-up questions, consider possible diagnoses and record the encounter. That documentation may include a clinical note commonly known as a SOAP note organised around the patient’s complaints, the doctor’s assessment and a plan. It is an essential part of medical practice, but it also takes time.
For a doctor seeing dozens of patients in a day, the burden accumulates. Every consultation generates information that must be entered into a computer or written down.
The problem is particularly acute in Nigeria, where the country’s low doctor-to-population ratio means that doctors are often overworked. One 2018 report found that doctors in Nigerian public hospitals could see more than 80 patients in a day. Intron’s research put that at 50 to 70 patients in a day.
In such an environment, improving how quickly doctors complete routine administrative work could make a meaningful difference.
It was this challenge that led Tobi Olatunji, a Nigerian-trained medical doctor who later worked at Amazon and Enlitic as a machine learning and natural language processing scientist, to explore a technology solution.
When he set out to solve a problem in the country’s healthcare system, electronic medical records (EMRs) seemed like a logical place to start. Digitising patient records could improve how hospitals store, retrieve and manage clinical information.
But as his team began working with healthcare providers, it discovered that doctors were still spending too much time entering information and wanted a voice-based solution.
If doctors could speak their observations and have them converted into text, they could spend less time documenting consultations and more time attending to patients.
There was, however, a catch: the voice recognition technologies available at the time did not adequately account for the way many Africans speak. That discovery prompted a change in direction.
Instead of concentrating on an end-user medical records product, Intron began developing the voice technology that other software products could use.
Read more: Talksign-1: A Nigerian is racing Google to build AI for deaf communication
Building the infrastructure layer for African languages
Speech recognition has become increasingly sophisticated, enabling people to dictate messages, transcribe meetings and interact with digital assistants.
Yet the quality of these systems depends heavily on the languages, accents and speaking patterns represented in their training data.
A model that performs well on standardised English recordings may struggle with a Nigerian English accent, Nigerian Pidgin or a conversation that moves between English and Yoruba. It may also have difficulty with specialised terminology and the different ways people pronounce the same words.
When a voice system fails to understand its users, the promised efficiency disappears. People must repeat themselves, correct transcripts or abandon the technology altogether.
For Intron, the answer was to develop proprietary models around the speech patterns it encountered in its target markets.
Temitope Ekundayo, who discussed the company’s development with me, described the gap the team encountered when it looked for existing voice models suitable for African use cases.
“It would surprise you to know that no major AI voice models had something that works for Africa,” he said.
Sahara voice AI models provide speech recognition and related capabilities that developers and enterprises can integrate into their own products.
The same infrastructure now serves several industries. A hospital might use it to transcribe clinical consultations; a law court could use it to document proceedings; and a financial services company could deploy it in a call centre or voice-based customer service application.
The expensive work behind building a voice model
Long before a model is trained, building a reliable speech model begins with data.
Voice AI systems need recordings of people speaking and accurate transcripts that help developers understand what was said. Depending on the application, the data may also need information about speakers, language variants, background conditions and specialised vocabulary.
The challenge is that much of the available speech data does not adequately represent everyday African speech. Studio recordings and people reading prepared sentences can help establish a baseline, but they do not capture the full complexity of spontaneous conversation.
Ekundayo notes that developing a model for a language can often cost more than $100,000, depending on the work required.
That cost can include collecting recordings, transcribing and annotating them, reviewing the results and conducting quality assurance before the data is suitable for training.
Intron has built a data collection platform called Zara Contribute to help source the voices it needs.
According to the company, its contributor network spans more than 70,000 people. The company has also described its approach as designed to comply with Nigeria’s data protection requirements.
After collecting the recordings, they must also be transcribed correctly, checked for errors and organised in a way that supports the intended use. Poor-quality labels can teach a model the wrong associations, while unrepresentative recordings can leave it unreliable for the very people it is supposed to serve.
According to Ekundayo, Intron prioritises languages with substantial usage and clear demand, including Afrikaans in South Africa, Swahili in East Africa and Hausa, Igbo and Yoruba in Nigeria.
Beyond prioritising languages with substantial usage and clear demand, it continues to expand the catalogue of languages it supports.
This year, it launched Sahara v2.5, which introduced 12 bilingual models and a trilingual model that supports Kinyarwanda, English and French within the same conversation.
“Version 2 understood how Africans speak, can hear, and can provide accurate numbers. Version 2.5 understands that Africans will switch between English and Yoruba in the same sentence and transcribes it,” he shared.
Read more: Emerging AI startups in Nigeria to look out for
Building for the realities of African businesses
Intron has had to account for infrastructure constraints and the different ways organisations communicate with customers.
According to Ekundayo, some telecommunications clients cannot easily integrate voice services through conventional application programming interfaces (APIs). In those cases, the company has had to work with Session Initiation Protocol (SIP) trunking, a technology used to connect voice calls to communications systems.
Other customers rely on WhatsApp as a primary channel of interaction, requiring solutions designed around that environment.
Pricing presents another challenge. Intron offers a credit-based approach for relatively straightforward uses such as transcription, while more complex enterprise deployments can require customised arrangements.
Ekundayo said the company’s goal is to keep the technology affordable and adjust pricing to the customer’s needs.
Intron says its technology is already being used across healthcare, financial services, and other sectors, including by organisations such as Helium Health, Mobi Health and eHealth Africa, as well as in judicial settings where the Ogun State Judiciary is one client.
The business of making voice AI useful
Intron’s trajectory reflects an opportunity for local AI startups. The largest technology companies already invest enormous resources in general-purpose models, but there is still room for specialised providers that understand the data, infrastructure and deployment requirements of emerging markets.
That does not mean the opportunity is guaranteed. Voice AI providers, for example, must compete on recognition accuracy, reliability, integration costs and data protection.
They must also demonstrate that their systems can handle sensitive information responsibly, especially in healthcare, finance, and legal settings. Eventually, they may have to compete with much larger and often better funded voice AI providers.
Intron says it does not train its models on client data, restricting itself to the data its contributors provide. If maintained in practice, that policy could be important for organisations concerned about the confidentiality of patient records, financial conversations and legal proceedings.
For Intron, the opportunity lies in turning African voices and languages into infrastructure that businesses can use. Its shift from medical records software to voice AI shows how a specific operational problem can expose a much larger market.
The original challenge was that doctors were spending too much time typing. The deeper issue was that the technology intended to help them was not adequately designed around the way they spoke.
That is what makes the 22-words-per-minute figure more than a statistic about keyboard skills. It points to a broader question about the assumptions built into digital products: who are they designed for, which languages do they understand, and what happens when their users do not fit the default?