Saheed Azeez Releases YarnGPT, a Text-To-Speech Model for Nigerians

Some months back, Saheed Azeez announced the release of Naija web, a Dataset of 230 Million GPT-2 Tokens from Nairaland that could improve the quality of data Nigerians have access to. We have written a story about it here. Now, he’s back again with YarnGPT. The young champ literally has a lot to serve Nigerians.

For us, the best of technology are those that solve a locally specific problem. So, what problem does YarnGPT solve?

As one of the pioneer text-to-speech models in Nigeria, YarnGPT—if judiciously harnessed, can enhance digital accessibility amongst Nigerians. The model can assist in reading out texts, websites, books, and documents in Nigerian accents and languages, thereby assisting Nigerians with visual impairment and reading disabilities like dyslexia. But, that’s only a tip of the iceberg.

Saheed Azeez is a final year Mechanical Engineering student from the University of Lagos, Nigeria. According to the release video on his twitter, Yarn GPT is a family of two TTS models used to basically convert texts to speech. One from English to Nigerian-accented English and the other for Yoruba, Igbo and Hausa.

His previous brainchild, Naija web, formed a huge part of what spurred Saheed to create YarnGPT. Naija web served as a source of experience and resilience-building and most importantly, it served as a source of motivation for the young lad.

“The Naija web project gained significant traction, which motivated me to do more. Today, I’m excited to announce YarnGPT.”

To develop the two YarnGPT models, Saheed started with a small-sized language model, SmolLM2, that was originally trained and developed by Hugging Face to understand and generate text. SmolLM2 lacked speech capabilities though.

Saheed extended and fine-tuned SmolLM2 with audio data and speech synthesis components. He incorporated audio tokens which enabled the model to convert text into speech. After which he began gathering a diverse Nigerian audio dataset, including dialogues from Nigerian movies, conversations and other local content.

Talk about the unseen efforts of machine-learning engineers!

After about two months of constant training on Google collab, YarnGPT had learnt Nigerian accents, intonations, and pronunciation nuances. Now, the model is not just able to produce natural-sounding Nigerian speech, but can also generate speech in local languages like Igbo, Hausa and Yoruba.

So far, the results of the YarnGPT TTS model have been impressive. And Saheed Azeez is excited about the possibilities that could ensue from it, while being open to improving the model. In fact, he has made the model open-source for developers to use and, for those who have the wherewithal, build applications around it. (You can explore YarnGPT here)

This development could help power chatbots, virtual assistants (like Siri, Alexa, or Google Assistant) in Nigerian accents and languages. It can be used to develop language learning apps for Nigerians trying to learn their native languages. From bridging the digital divide to Edtech, who knows how far-reaching YarnGPT can be?

Let’s not even talk about what wonders our Nigerian Youtubers and TikTokers would make out of it once they get hold of it.

“I have been gathering data and learning how to build audio-based models over the last two months. And thanks to God, I have been able to come up with something!”