WILLIAM Gibson’s quip ‘The future is already here — it’s just not evenly distributed’ never feels so apropos as when travelling in sub-Saharan Africa. In a tech hub like Addis Ababa, it’s a short walk from a modern office teeming with programmers collaborating online with overseas colleagues to a ghetto neighbourhood of crammed-together shacks without electricity or clean water. And then only a couple hours’ drive along crowded, dusty, potholed roads to rural villages where swarms of industrious kids buzz around outdoor marketplaces helping do business during what are supposed to be school hours.
Technology’s power to overcome divergences in politics, history and culture often operates with a significant time lag, and the impact of advanced tech on the developing world has proved complex and erratic.
There’s a long list of viable applications with well-known potential to help salvage and supercharge African economics. Credit scoring for unbanked individuals and small/medium enterprises. Agile, adaptive supply chain management — including for supply chains with key segments that are largely off-the-books and artisanal. Smart power for grids involving components of widely varying age, quality and instrumentation. Educational software that accounts for local language, culture and knowledge. Automated tools for diagnosing health issues in humans or agricultural crops — and helping discover cures.
I have become convinced, however, that there’s one AI application with more profound potential than all the others to provide transformative value to sub-Saharan Africa and help bootstrap it into the upper reaches of the global economy: speech-to-speech machine translation for under-resourced languages. Specifically, speech-to-speech machine translation for languages with no writing systems and only a modest amount of audio recordings to use for training.
There are more than 2,000 distinct languages in Africa and many more dialects. This is more than a third of the world’s languages, in a continent with less than a seventh of the global population. Around 80 percent of these languages have no written form. Many Africans speak multiple languages, but if none of these is a language featuring a large body of written information, this still doesn’t help in terms of accessing the corpus of human knowledge.
Mobile phone penetration in Africa is nearing 50 percent. In 2019, 477 million Africans had mobile phones, around 45% of the total population, and nearly 60 percent of these were smartphones providing internet access. Around two-thirds of the African population lives in areas covered by mobile internet signals. All these numbers are increasing fast; the number of mobile internet subscribers in Sub-Saharan Africa has quadrupled since the start of the decade.
Extrapolating these trends, we can see that within a few years, nearly all Africans could potentially have mobile internet access, but many will not be able to use the internet to its full potential because of linguistic incompatibilities.
Illiteracy is also a major issue, with around 35 percent of adults in sub-Saharan Africa functionally illiterate. However, the impact of illiteracy on internet usage can be palliated by the adoption and adaptation of speech interfaces to various online services. If a person doesn’t know any of the languages commonly used online, and there’s no available machine translation from the languages they know to the ones that are useful on the internet, there are few ways to salvage the situation.
Fortunately, speech-to-speech machine translation is advancing rapidly, with recent papers from China, Taiwan and other places breaking new ground via automatically decomposing speech into phonemic symbols and using this breakdown to help guide the neural machine translation process. Accuracy increases each year and is now just barely at the level of producing marginally comprehensible translations. It would seem that, if progress continues apace, within two to five years we may have systems that can be trained to translate the vast majority of spoken African languages into English, Chinese or any other language of choice.
Of course, it will be necessary to actually take the time to apply this software to recordings of various African languages, but the Masakhane project, spanning numerous African nations, has pulled together a community of African technologists who are enthusiastic to do precisely this. The Masakhane project members fully understand the main point I’m making here: Effective machine translation for the full scope of African language has tremendous potential both humanistically and economically.
Assuming this happens, then the next question becomes: Who pays for all the processing required to supply this translation to the billion people populating the African continent? Neural machine translation models tend to be large — very expensive to train and moderately expensive to deploy and query. Big tech companies have little short-term motivation to provide such services to large numbers of people with minimal purchasing power, sometimes living in economies that limit the exchange of local for global currencies. Providing voice interfaces to online resources is something big tech is pursuing avidly already, for reasons unrelated to Africa. Speech-to-speech MT is more critically useful in the developing world where literacy rates are lower and there are more commonly utilized languages without written form or with minimal written resources.
One possible solution is to distribute the processing power needed for querying neural ML models and updating them according to new data gathered from users. Projects like GOLEM Network and our company’s new spinoff NuNet.io, which is focused specifically on AI applications, aim to distribute complex software processes among multiple possibly weak processors, like those running on thousands or millions of peoples’ mobile phones. If one has multiple processors (e.g., smartphones) in a local area like a village, but internet connectivity is costly or erratic, one can use mesh network tools like LibreMesh to locally pull coordinates of the various processes to allow them to support a common computational process like neural machine translation.
We can see here a very definite and concrete path via which some of the most advanced technologies on the planet today may soon start helping some of the most disenfranchised and impoverished people in the world to encounter the whole scope of human knowledge and enter fully into the world economy. The pieces are falling into place, but as always, it will still take imagination, dedication and coordination to make it happen.