How Google aims to teach AI Africa’s 2,000 languages

Voice as Africa’s digital entry point

When Abdoulaye Diack, program manager at Google Research, explains the origin of WAXAL, he begins with a single word.

WAXAL means “speaking” in Wolof, a language widely spoken in the Senegambia region. The name reflects a central idea. In Africa, where more than 2,000 languages are spoken and many are primarily oral, voice is often the main gateway to technology.

For years, digital systems have focused on text. In much of Africa, however, language lives in conversation across markets, farms, schools, and clinics. AI systems that cannot recognise accents, intonation, or code switching struggle to serve large parts of the population.

WAXAL, developed by Google Research Africa, aims to address that gap by building foundational speech data for African languages.

The data imbalance problem

More than half of all websites operate in English and a few Western languages. African languages barely appear in global digital datasets. Many lack extensive written records. Some do not have standardised spelling systems.

Modern AI models rely heavily on large volumes of digital text. When that text does not exist, AI systems cannot learn effectively.

Diack says researchers have long recognised this structural disadvantage. Without sufficient data, AI systems mishear, mistranslate, or ignore entire communities.

For example, voice tools often struggle with francophone African accents or regional dialects. The technology exists, but it lacks local tuning.

Building a large speech foundation

After three years of development, WAXAL launched officially in February 2026. The initiative produced more than 11,000 hours of recorded speech across 21 Sub Saharan African languages, including Hausa, Yoruba, Luganda, and Acholi.

The project collected nearly two million individual recordings. It also invested in more than 20 hours of studio quality recordings to create more natural synthetic voices for digital assistants.

Universities such as Makerere University and the University of Ghana led much of the data collection. Local partners retain ownership of the datasets. Google released them under open licences that allow commercial use.

Diack says Google provided guidance and funding, but partners control the data. Within days of release, developers downloaded the dataset more than 4,000 times.

Why translation alone is not enough

Google already supports translation across many languages. However, translation depends on parallel text, which many African languages lack.

Voice AI must go beyond text. Many people interact with technology through speech rather than typing. Voice systems must handle dialects, slang, mixed languages, and atypical speech patterns.

In Ghana, the UGSpeechData initiative generated over 5,000 hours of audio. That effort helped develop a maternal health chatbot in local languages and expanded support for users with atypical speech, including stroke survivors and deaf communities.

AI systems often fail to recognise such speech variations. Diack says adapting to diverse speech patterns is essential for inclusion.

A growing ecosystem of African AI

Google is not alone in this effort.

Masakhane, an open research collective, has developed translation tools in more than 45 African languages. South Africa’s Lelapa AI builds commercial language processing tools that reflect dialects and urban speech patterns. Ethiopia’s Lesan AI focuses on culturally nuanced translation models.

Meanwhile, Meta runs its No Language Left Behind project, which translates across 200 languages, including many in Africa. Microsoft integrates African languages into its translation services and invests in agricultural AI tools.

The approaches differ. Some focus on scale. Others prioritise depth and cultural precision. Google’s strategy emphasises speech data and ecosystem building.

Balancing collaboration and sovereignty

The involvement of global technology companies raises questions about data sovereignty. Critics ask whether large firms may create dependency by controlling foundational datasets.

Diack acknowledges these concerns but argues that collaboration accelerates progress. He stresses that open licensing allows developers to build independent products without relying solely on Google platforms.

Google has also released open weight translation models such as Translate Gemma, which developers can download and adapt independently.

Infrastructure remains critical

Voice AI depends on connectivity and computing power. Diack says AI development requires people, data, and infrastructure.

Google has invested in undersea cables such as Equiano to strengthen broadband capacity in several African markets. Reliable connectivity supports cloud services and local data centres, which are essential for AI research and deployment.

Africa’s young population offers demographic potential. However, without research investment and digital infrastructure, that potential may not translate into leadership in AI.

A long term ambition

WAXAL currently covers 27 languages, including several Nigerian languages such as Hausa and Yoruba. Other languages include Akan, Acholi, Swahili, Lingala, Malagasy, and Shona.

Teaching AI all 2,000 African languages remains a long term ambition. Diack describes it as a generational project.

He points to education, agriculture, and health as priority sectors. Voice AI could deliver weather forecasts, medical guidance, or farming advice in local languages.

Projects such as PlantVillage Nuru, developed through partnerships involving Pennsylvania State University and agricultural research institutions, show how solutions built in Africa can scale globally.

For Diack, success will mean seeing startups and researchers use African language data to create new services and publish research beyond English.

If speech recognition improves, billions of spoken words across the continent may finally become part of the digital ecosystem rather than remaining invisible.