Bridging the Digital Language Gap: The Evolution and Impact of the Khaya App

Introduction

In the fast-evolving landscape of artificial intelligence and machine learning, the Khaya App represents linguistic diversity and cultural preservation. Developed to address the gap in language translation tools for African languages, Khaya has been on a mission to empower local communities and protect cultural heritage in the digital space.

In this article, we explore the inception, challenges, and evolution of Khaya, shedding light on its contributions to communication and accessibility for speakers of Ghanaian languages and beyond.

Inception and Mission of Khaya App

The Co Founder of Khaya, Dr. Paul Azunre, was inspired by a critical observation: the lack of attention from major tech companies towards building tools for African languages. 

The mission of Khaya extends beyond mere translation. It aims to equip nations with the technological capital and capacity to retain control of their culture and heritage in the digital era. 

Recognizing the risks associated with disinformation and social engineering campaigns, Khaya emphasises the importance of building and owning tools locally.

“We looked around and did not see any of the big tech companies building tools for our languages. Google Translate for example, only had some of the bigger languages in Africa – in terms of the number of speakers – like Swahili. We had to start doing something because it felt like, as a culture and nation, we were falling behind on this.”

“Beside that, as the risks surrounding disinformation and other social engineering campaigns through more advanced AI tools become more apparent, it is important for every nation to retain full control of their culture and heritage in the digital space. This occurs by building technological capital and capacity by locals for locals. That is part of our mission,” Dr. Azure said.

Challenges in Development

Developing a translation app for Ghanaian languages presented a unique set of challenges. These challenges primarily revolved around the scarcity of data. Unlike more widely spoken languages, the lack of open data for languages like Twi meant building the translator from scratch. Building from scratch also meant requiring significant funding. However, funding challenges were made even bigger by the thought that languages with smaller speaker populations were less lucrative for investment.

“The typical thing researchers face right away in this space is the lack of data for our languages. No open data existed to create a Twi translator, so we had to first build it from scratch. To do this well requires significant funding (several thousand dollars per language), and so a big part of the challenge is figuring out how to come up with that money.”

“Writing proposals for funding for us has been a frustrating experience; it opens you up to exploitation and idea theft, and frankly, it seems most people don’t care about languages in our country because the number of speakers (and therefore the potential number of users for any tool that is built) is smaller than a lot of languages in, say, Nigeria or East Africa. I have seen our entire culture systematically frozen out of progress and opportunities due to this factor. So we have to stay creative and keep looking for alternatives, mostly relying on ourselves.”

Evolution and Expansion

Since its initial release, Khaya has expanded its language repertoire, adding Northern Ghanaian languages like Dagbani and Gurene. The app’s collaboration with Harvard University further extended its reach to Kenyan languages like Kikuyu, Kimeru, and Luo. 

Now, Khaya features the integration of Automatic Speech Recognition (ASR), enabling users to interact with the app through spoken language.

“We started with Twi, Ga and Ewe and then expanded into Northern Ghana with the addition of Dagbani and Gurene. We also added Yoruba because we could find data we could use right away and it seemed there was still a need to fulfil there. We have also collaborated with a group at Harvard University – the African Language School there – to add Kenyan languages Kikuyu, Kimeru and Luo. These are some of the biggest native languages in Kenya, and ours was the first solution to address them. How insane is that? We have also added speech recognition in addition to the initial text translators over time. Soon we will be introducing Text to Speech to enable applications such as “Siri for Gurene” etc.”

“As you may know, the level of literacy in our communities isn’t very high – in particular when it comes to reading and writing our local languages properly. To be truly useful to everyone in our communities, the tools need to be able to accurately detect and transcribe spoken speech, as well as to respond using a natural human voice. Speech recognition is the transcription part. We will introduce text to speech very soon to handle the response part. It is a work in progress.”

Conversational Capabilities and Challenges

Khaya’s future includes incorporating features similar to Chat GPT for conversational engagement. This aspect is still in the early development phase. The challenges lie in adapting existing models, which are based on English Language Models (LLMs), to local languages. Fine-tuning these models directly in local languages demands substantial investment, presenting a multifaceted problem with no overnight solutions, according to Dr. Azunre.

“We have not yet released this in our product. It is currently in the early development phase and we are still exploring to which extent various approaches to solving this problem are currently effective. Sometimes it does impressive things in testing, but that does not mean it is yet ready to be released to the public. We are actively working on it though, so sometime next year we expect to be able to release a version for some languages. The cultural nuance is one of several factors that necessitate care. Currently the models we have are based on translating English LLMs, and this is of course not ideal. We are exploring fine-tuning them in local languages directly, but that is a resource problem requiring significant investment to execute properly – investment we presently do not have. So as you can see it is a multifaceted problem with no overnight solutions.”

Accuracy, Sensitivity, and Community Involvement

Ensuring accurate and culturally sensitive translations is paramount for Khaya. The key lies in involving the communities it serves throughout the development process. Real humans from these communities contribute to data collection, testing, and the development of UI and algorithms. This community-centric approach ensures that the tool is well-suited to the linguistic and cultural nuances of African communities.

“At the end of the day, the key is to involve the communities that are going to be served by these tools in the entire development process. Real humans from these communities must be involved in collecting the appropriate data to ensure accurate representation, well-balanced for gender, etc. They should also be involved in testing, of course, but also in the development of the UI and the underlying algorithms. Because this will ensure the entire tool is well thought-through for the African community. For example, I feel the constraints faced by African researchers would influence our research to focus more on smaller, more efficient models than a company like Meta would prefer to focus on. Or prioritise offline use, as another example. We have our own path to addressing every aspect of this problem that we should explore.”

Positive Impact and Success Stories

Khaya has garnered positive feedback and success stories from users. Many in Ghana use the app to enhance their English language skills, especially through the speech recognition feature. The app’s utility in education extends to helping individuals learn languages in both directions. In the diaspora, Khaya has facilitated cultural connection by aiding children in learning local languages.

Future Plans and Vision

Khaya’s future plans include expanding its coverage to encompass the entire Ghanaian linguistic map. Collaborations with groups like Lesan AI focusing on the Horn of Africa demonstrate a broader vision to make language technology accessible and beneficial across the African continent. Continuous improvement and the introduction of new features, such as text-to-speech, are on Khaya’s agenda, turning the app into a dynamic and evolving tool.

“We hope to eventually be able to cover the entire Ghanaian linguistic map. Through collaborations with similar groups in other parts of Africa, e.g., Lesan AI focusing on the Horn of Africa, we hope to also help in whatever way we can to ensure this technology reaches and helps every part of Africa. Beyond that of course, the translators, speech recognition systems, chatbots and text-to-speech systems all need to continue improving, becoming more efficient, staying updated with new linguistic trends, etc. Think of it as a living, breathing thing that we will need to continue investing in. And of course, new exciting applications will continue to be discovered.”

Technology as a Guardian of Culture

In conclusion, Khaya embodies the power of technology in preserving and promoting African languages and cultures. In a highly digitised future, language tools become essential for preventing cultural extinction. By making these tools available, Khaya ensures that future generations can not only learn the languages but also reconstruct and understand the nuances of African heritage, creating a digital canvas for the rich tapestry of cultures across the continent.

About Dr. Paul Azunre

Dr. Paul Azunre is a Ghanaian scientist who has made significant contributions to the fields of Machine Learning, Natural Language Processing, and Optimization. He holds a PhD in Computer Science from MIT and has served as a Principal Investigator on several DARPA research programs. Dr. Azunre founded Algorine Inc., a research lab dedicated to advancing AI/ML and identifying scenarios where they can have a significant social impact. He is also the author of the newly released book “Transfer Learning for Natural Language Processing,” which summarizes some recent related NLP model architectures, such as BERT, GPT, and ELMo.

Dr. Azunre is a former captain of the Opoku Ware School team that won the 2002 edition of the National Science and Maths Quiz, a popular nationally-televised math and science show in Ghana.