• D.C.
  • BXL
  • Lagos
  • Riyadh
  • Beijing
  • SG
  • D.C.
  • BXL
  • Lagos
Semafor Logo
  • Riyadh
  • Beijing
  • SG


Orange partners with Meta and OpenAI for African languages AI models

Nov 27, 2024, 10:57am EST
africaAfrica
Open AI logo on a computer screen
Ishmael Daro/Flickr
PostEmailWhatsapp
Title icon

The News

LAGOS — French telecoms company Orange plans to partner with Meta and OpenAI to develop new artificial intelligence models trained on African languages.

Orange, a major mobile network and internet provider in French-speaking Africa, said the project will create models that allow customers “to communicate naturally in their local languages” with its sales and support teams. The company said it will offer the developed models for third-party health and education uses through a free license.

Wolof and Pulaar — two languages spoken by about 22 million people in and around Senegal and Gambia in West Africa — will be the first to be considered for the project, Orange said on Tuesday. The company’s long-term goal is to expand AI’s capacity in local languages spoken where it operates in Africa and the Middle East.

AD

The partnership is due to begin next year.

Title icon

Know More

Orange’s project will look to “fine-tune” existing models by OpenAI and Meta.

OpenAI’s ‘Whisper’ automatic speech-to-text model was rolled out in 2022. About a third of the audio dataset it was trained on consisted of non-English data, according to OpenAI. Meta’s ‘Llama’ text model powers the AI chatbots found in the company’s apps, including Facebook Messenger and WhatsApp.

Orange hopes that adapting these large language models will enhance access for “illiterate populations, who are currently unable to benefit from the potential of artificial intelligence.”

Title icon

Alexander’s view

Orange’s project comes amidst rising interest in making AI tools available to African audiences. Multiple efforts are underway but it may not be an easy undertaking for a diverse continent in which more than 3,000 languages are spoken.

AD

In August, South African startup Lelapa rolled out a multilingual large language model. The company’s InkubaLM is designed to translate and transcribe African languages, beginning with Swahili, Yoruba, IsiXhosa, Hausa, and isiZulu. Kenya-based Jacaranda Health, which is tailored towards caring for expectant mothers, also recently updated its LLM in the same languages. Five years ago, two Nigerian software developers made an original attempt at creating a model that translated pidgin English spoken in the country.

In Africa, telecom companies are among the largest companies and maybe the largest platforms through which consumers interact with each other, making them effective vehicles for mainstreaming AI use. Orange’s announcement that this project is aimed at improving customer satisfaction clearly shows there is a financial incentive for the undertaking. But there is a potentially wider social benefit in their intention to freely license the models for public health and education.

Africa’s biggest AI success story so far is a commercial one, following last year’s $680 million acquisition of Tunisian AI startup InstaDeep by German company BioNTech. Maybe Orange’s LLM project will be a major step towards creating “a blueprint for how AI can be used to benefit those currently excluded,” as the company says.

AD
Title icon

Room for Disagreement

Orange has not disclosed its planned approach to a crucial aspect of creating an LLM — how it will collect the volumes of African language data required to train the new models.

OpenAI’s use of Kenyan contractors to train ChatGPT, its flagship product, has become infamous for what workers have described as exploitative, low pay work that was often toxic.

AD
AD