TIBaLLi: Making AI Work for Internet Inclusiveness in the Global South

Research Question

1. Can advanced AI methods (ML, NLP) be reconstructed so as to make the Internet more inclusive for communities in low-resource environments in the Global South — such that local weather data combined with Internet-based global climate information become sharable beyond the Internet’s current boundaries and in people’s own language?

Why Is This Research Important?

There is a digital divide in terms of digital information, not just between the Global North and the Global South but also internally in most developing countries. This divide is partially due to inequalities resulting from the initial methodologies that guided the development of the Internet. A decentralized Internet can tackle the major hurdles of the digital divide from a bottom-up approach by focusing on voice technologies, local under-resourced languages, and existing communication technologies.  

This research tackles the issue of the digital divide by providing tools that build a decentralized, context-aware, community-centred but still global Internet; combine knowledge from communities, institutions, organizations and research to realize the aspects of the colonized Internet that can be circumvented by available or new technologies; and results in methodologies and systems that foster easier access to digital information in low-resource environments.  

Methodology

The overall methodology of this research is Participatory Action Research (PAR). The key feature that sets PAR apart from other methodological approaches is that individuals and communities are conceived of as first-class partners in a co-creation approach.   

Within the PAR paradigm, various kinds of Information Systems (IS) and Software Engineering (SE) components are designed and deployed, including Artificial Intelligence (AI) tools. The Design Science methodology for Information Systems and Software Engineering is followed to construct, test and evaluate ICT/IS artefacts in interaction with their real-world context. 

Concerning data gathering, the project gathers audio recordings of brief utterances from community members from which data analysis is made by AI methodology for Automatic Speech Recognition (ASR) for low-resource environments. It also gathers local content data regarding a specific but important domain, weather and climate information as locally relevant for farming, to build up a limited vocabulary out of which locally useful farmer messages concerning weather and climate can be generated. This provides a basis to integrate Internet-based global climate information with local data and local ways of expressing weather information.  

The quality, relevance and adequacy of such weather/climate information messages for communities are evaluated through interviews and community workshops with round-table conversations. 

Media and Publications

• Tiballi 
• Resourcing Small Indigenous Languages in the Field This paper presents a method for resourcing indigenous languages, with the aim to build targeted voice services based on an automatic speech recognition system. 
• From Raindrops to Insights: AI Empowers Hyperlocal Rain tracking in Ghana This paper illustrates the development and application of an AI-based system aimed at addressing the critical need for accurate and accessible meteorological data among farmers in Northern Ghana
• Telegram bot  The test bot expects a voice message from the user, containing one of the words in the vocabulary. This message is then processed by the ASR model, which results in a written version of what the user said. This text is then converted back to speech (using the slot-and-filler method) and sent as a voice-message back to the user. The user can then listen to the response of the bot, and confirm whether the result was correct or not.
• Rural Development and the ICT4D Plug-In Principle for Information and Communication Technologies This paper formulates the ICT4D Plug-in Principle. Its two keypoints are: (i) collaborative; (ii) contextual.
• Bridging the Digital Divide This chapter discusses the disparity in access to digital services, which exists between countries, regions, communities, and people in the world.
• Integrating Indigenous and Scientific Knowledge for Sustainable Food Systems in Africa The Plug-In Principle advocates that advancements in science and technology should enhance rather than replace existing indigenous knowledge.
• Recordings of spoken Dagbani for the letters zero to ten and the words yes and no The dataset contains recordings of the words zero to ten, and yes and no, spoken in Dagbani. The recordings have been made by the Tingoli and Nyankpala communities of the northern region of Ghana.
• Energy-efficient phonetic matching for spoken words of rare languages in low-resource environments This thesis introduces a novel method of handling spoken rare languages, by introducing a method to automate matching of their spoken words, and therefore attempts to help solve a key problem in low-resource environments.
• Speechmatching Python Package The speechmatching package allows for calculating similarity scores between short audio fragments containing speech. It then uses these similarity scores to find matches over multiple groups of recordings.