Who Gets Heard by AI? Nigeria’s Language Gap in Voice Technology.

By Chisom Okafor

Ask Siri a question in Yoruba. Speak a command to Alexa in Hausa. Try to hold a conversation with ChatGPT’s voice mode in Igbo. In most cases, the result is the same.

You get silence, misrecognition, or a response in English that misses the point.

This is a major technology challenge the Big Techs are dealing with now, even startups that rely on open-source API are not exonerated. About 175 million Nigerians speak Hausa, Yoruba, and Igbo as their primary languages.

Tens of millions more use Nigerian Pidgin, Efik, Tiv, Kanuri, and many other indigenous languages every day.

What they face is a structural exclusion from one of the fastest-growing areas of technology.

Global voice AI revenue reached about $8.5 billion in 2024. Systems such as Amazon’s Alexa, Apple’s Siri, Google Assistant, and ChatGPT’s Advanced Voice Mode have become key interfaces.

People use them to access information, control devices, interact with services, and manage daily tasks. Yet none of these systems work meaningfully in Nigerian languages.

The reason is not mysterious. It is data.

“These systems learn from training data,” said Kaosarat Aina, a researcher in Linguistics at Indiana University Bloomington who focuses on computational processing of African tonal speech. “Most training data is in English, with some Mandarin and Spanish.

Nigerian languages are mostly absent. A system cannot learn what it has never been exposed to.”

Aina has spent years advocating and building a phonemically and tonally annotated speech corpus for Yoruba learner language. She spoke as part of this investigation into why voice technologies remain inaccessible to most of Nigeria’s 220 million people.

Nigeria’s major languages present a deeper challenge. Yoruba, Igbo, and to some extent Hausa are tonal languages. Pitch determines meaning. In Yoruba, the same sounds can form different words depending on pitch. This is not a rare feature. It is central to the language.

Aina noted that voice recognition systems designed for English are not built for this. English relies more on consonants, vowels, and word order. Tonal languages require a different approach.

“This is not solved by adding more general audio,” Aina explained. “You need speech data with phonological annotation. That means recordings labeled with tone and structure. This requires linguists who understand the language. It is not just about collecting audio. The corpus comes first. Everything else depends on it.”

The lack of such corpus data is foundational to why Nigerian languages remain excluded from effective voice AI. Researchers say this is also the clearest path forward.

On September 20, 2025, at the 80th United Nations General Assembly in New York, Nigeria’s Minister of Communications, Innovation and Digital Economy, Dr. Bosun Tijani, announced N-ATLAS. This stands for Nigerian Atlas for Languages and AI at Scale. It is an open-source multilingual language model supporting Yoruba, Hausa, Igbo, and Nigerian-accented English.

The model was developed by the National Centre for Artificial Intelligence and Robotics with Awarri Technologies, a Lagos-based company. It was trained on over 400 million tokens of multilingual data. It represents a major step toward building AI systems for Nigerian speech.

“N-ATLAS is more than a language model,” Tijani said. “It is a national commitment to inclusion and global contribution.”

The ambition is clear. Aina agrees that this is only a starting point.

“Releasing models like this is the starting point to solve the data problem,” Aina said. “A model reflects its training data. If the data lacks proper tonal detail, the system will also lack accuracy. The announcements are encouraging, but the real work is in building the datasets.”

Engineers involved in the project share this view. Sunday Afariogun, lead project engineer at Awarri, described N-ATLAS v1 as a foundation for developers, not a finished product.

N-ATLAS is not the only recent effort. Intron, a Nigerian-founded AI company, has developed speech recognition systems for African contexts. These systems use real-world audio instead of studio recordings. The data comes from environments such as clinics, call centers, and other everyday settings.

This helps models handle real usage conditions.
“Recording speech is the first step,” Aina said. “The real challenge is annotation. You must capture the structure of the language accurately. That is where the bottleneck is and where most of my works are addressing.”

Voice AI is now critical, they are used in healthcare, education, customer service, and government systems.

As these systems expand, the languages they support will determine who benefits and who is left out.
Nigeria has over 500 languages. Even the three targeted by N-ATLAS represent only a portion of the population. Getting these languages right is essential. It requires building annotated speech corpora that reflects tonal structure and real usage.

This is not just a technical task. It is a foundation for Nigeria’s broader AI ambitions.
“The model places Africa’s voices at the center of AI,” Tijani said.

Researchers working in this field understand what it will take to make that goal a reality.