One, Two…Twi! Voice dataset reaches 100 languages

Common Voice, a project to help make voice recognition open and accessible to everyone, has just added its 100th language, with the latest addition arriving from West Africa.

Twi, a language that is native and bilingual to approximately 18-million people across Ghana, Benin, and the south-east regions of the Ivory Coast in Western Africa, has been added to the the Common Voice V10.0 database.

While voice recognition technology is revolutionising the way we interact with machines, currently available systems are expensive and proprietary, needing payments and licences. The Mozilla Foundation started the Common Voice  initiative to make voice recognition technologies better and more accessible for everyone. 

Mozilla Common Voice is a massive global database of donated voices that lets anyone quickly and easily train voice-enabled apps in potentially every language.It is collecting voice samples  in widely spoken languages as well as those with smaller populations of speakers.

The Mozilla Foundation says that publishing a diverse dataset of voices will empower developers, entrepreneurs, and communities to address this gap themselves.

“The heart of this project is making it easier for language communities around the world to tap into the possibilities of speech technology —

creating a healthier and more open AI ecosystem,” says EM Lewis-Jong, Common Voice product lead.

According to the State of Internet’s Languages Report, the insignificant representation of African languages online continues to reinforce a form of colonial imperialism. The report, produced by Whose Knowledge?, Oxford Internet Institute, and the Centre for Internet and Society, states: ‘’The vast majority of African languages are not supported as an interface language by any of the platforms we surveyed, and as a result, more than 90% of Africans need to switch to a second language in order to use the platform — which for many will mean a European-colonial language.”

While African languages account for at least a third of spoken languages worldwide, only a handful of products support these languages — despite the majority of these languages existing in oral forms rather than in written text.

Changing this trajectory is a pool of motivated language community builders like Daniel Agyeman, a contributor to Common Voice. He says this endeavour is one that reconnects him with his culture: 

“I was born and raised in the UK but I am of Ghanaian descent. As a Ghanaian living in the diaspora I am drawn to activities that will help me to connect with my home country and specifically improve my Twi speaking skills. Currently, I have not come across any usage of Twi in voice technology, therefore I was very excited at the prospect of creating a Twi voice dataset which can be used to create the first ever speech recognition system in Twi.’’

Daniel has been engaging family, friends, and other Ghanaians living in the diaspora to collect sentences in Twi and upload them on the Common Voice platform. So far, the Twi language community has gathered over 40,000 text sentences and is currently inviting native or bilingual speakers to contribute to the initiative by donating their voices and validating other speakers’ contributions.

Through the support of the Gates Foundation, NVIDIA, and GIZ, the 11th Common Voice dataset release is set to exceed 23,000 hours. This community-driven dataset is grown through community mobilisation with the support of over 400,000 volunteers around the

world.