Saheed Azeez, a Mechanical Engineering student at the University of Lagos, has created YarnGPT, a text-to-speech AI model that can read English and Nigerian languages in a local accent.
Azeez first gained attention in November 2024 with Naijaweb, a dataset of 230 million GPT-2 tokens based on Nairaland. Now, he has taken on a more ambitious project—building an AI voice model that captures the nuances of Nigerian speech.
Speaking to TechAfrica, Azeez described the project as a tough challenge, especially in gathering quality data.
Data challenges and innovation
To build YarnGPT, Azeez needed a large dataset of Nigerian voices. He extracted audio from Nollywood movies but found that many lacked high-quality subtitles. To compensate, he used open-source datasets from Hugging Face and leveraged Oute AI, a platform specializing in text-to-speech models.
“Replicating what has been built overseas isn’t that hard, but data always gets in the way,” Azeez said in an interview with TechAfrica.
Training on a budget
Without access to a GPU, Azeez trained YarnGPT on Google Colab, spending $50 (₦80,000)—only for the model to fail. Undeterred, he refined his approach and spent another $50 to train a better version, a process that took three days.
The AI model relies on tokenisation, converting both text and audio into numerical representations. “Unlike text, which has clear breaks between words, audio is continuous, so it needs to be broken into smaller pieces the model can process,” Azeez explained to TechAfrica.
Viral success
Azeez showcased YarnGPT in a two-minute video, which attracted 138,000 views on X (formerly Twitter) and caught the attention of Timi Ajiboye, co-founder of Hellicarrier.
To make the video, Azeez and his team rearranged a friend’s house—much to the displeasure of the homeowner’s mother. But the results paid off, bringing widespread attention to the project.
A step forward for AI in Nigeria
YarnGPT has multiple applications, from content creation to accessibility for non-English speakers. It can pronounce English in a Nigerian accent and read in Yoruba, Igbo, and Hausa.
Despite his achievements, Azeez acknowledges Nigeria’s AI limitations. “We’re not even in the race compared to the U.S. and China. But we can localise AI for our own needs instead of building from scratch,” he said.
With the Nigerian government pushing for AI development, innovators like Azeez could help bridge the gap.
Reported by TechAfrica
