Understanding Africa's multitude of voices
While voice recognition technology is working elsewhere, Africa's many accents and dialects make success here more difficult.
There's a raft of applications available today where speech is the most efficient and convenient way of communicating, which is why speech or voice recognition technology has become so pervasive.
Research indicates half of all searches will be voice-based by 2020, and this, in turn, will likely lead to a whole new ecosystem of applications and interactions being developed.
In fact, the global speech and voice recognition market size is estimated to reach USD31.82 billion by 2025, according to a new report by Grand View Research. But, how much of that will benefit Africa and its growing tech hubs?
The challenge for those of us living in Africa, however, is that developers are focusing far less on languages from nations here and in Asia. This is a problem because, while the ability of computers to comprehend the human voice and language has improved significantly, there remains a divide between the First World, notably English-speaking countries like the UK and US, and those in the developing world, who use less widely spoken languages.
"There are companies and non-profit voice data collection sites around the world and we support these efforts to close the gap by building various global voice data sets," says Andrew Dawson, Managing Director of Future Fragment. However, as of yet there has been little visible and/or vastly successful drive to increase the share of African-based voices in these data sets.
Dawson suggests if we are to use voice recognition technology effectively in Africa, particularly in areas like voice-to-text conversion or the analysis of voice for customer experience (CX) purposes, we need to build a system that can accurately translate and understand African languages. However, the numerous African accents and tongues found across the continent have, until now, made this next to impossible.
"I have heard that some of the telco providers in SA have tried using voice recognition technologies, but have found they are receiving less than 40% positive results. Getting it right just two out of every five instances is simply not good enough, especially in regard to something like CX. What's needed is a way to significantly refine the voice recognition model," he says.
In order to do this, a large enough data set of continental languages and accents is required to run the models against, in order to improve accuracy. "What we believe is the answer is to craft a data set of purely African voices, enabling us to build a more accurate model. To this end, we are now utilising a unique crowd-sourcing approach that we hope will ultimately obtain around 100 000 voices from all across the continent."
Dawson indicates that, in order to get as many people as possible from all over Africa to contribute to this, the company has begun using social media to target communities that can, in turn, create viral growth across the region, rapidly and effectively.
"We have received a lot of support from the African AI community in getting our message out there, and we are also encouraging people to not only link to the company's social media accounts, but also to share these links with friends. By spreading the message via this three-degree approach, we hope to quickly achieve the 100 000 voices we require."
Being a part of the evolution of the African Voice (AVoX) space is easy and multi-modal; either download the Android application or go to the Web site, thereafter complete the steps that ask you for additional information such as region, language and ethnicity. "We require this information so that we can effectively train our model and provide accurate detection of specific dialects. This information is crucial to determining the differences and nuances in how words are pronounced and how accents impact and shape various dialects across regions," says Dawson.
While there are clearly commercial applications for such a project, other applications exist in training and learning institutions, not to mention data scientists and the greater AI community. "It's about significantly improving voice recognition technology across Africa, along with all its attendant benefits.
"I believe that in the next couple of years, voice recognition is going to become the fastest growing biometric option, so the results of this project will provide potentially the most accurate service of this type across the continent and should go a long way towards improving the state of voice recognition in Africa. We feel that by developing this technology, we will ultimately be able to pay back, in some small way, the various communities in Africa that have lent us their voices," concludes Dawson.