Sign in Subscribe
College faculty

African Languages for AI: The Project Gathering a Huge New Dataset

Strategic Insights

— 1 min read

African languages are crucial to the development of AI, as they hold culture, values, and local wisdom. However, the lack of language data and poor AI performance in African languages have denied millions of people access to global resources, news, and healthcare information. The African Next Voices project aims to bridge this gap by collecting speech data for automatic speech recognition (ASR) and developing a dataset that accurately reflects authentic language use in African communities.

Key Takeaways:

  • The African Next Voices project is collecting speech data for ASR in five languages in Kenya, Nigeria, and South Africa, with a focus on capturing diverse age groups, genders, and educational backgrounds.
  • The project aims to collect 100 hours of speech data and is prioritizing languages that are largely spoken, such as Nilotic, Cushitic, and Bantu languages in Kenya, and Bambara, Hausa, Igbo, Nigerian Pidgin, and Yoruba in Nigeria.
  • The dataset will be used to develop models for captioning local-language media, voice assistants for agriculture and health, and call-centre support in African languages.
  • The project is working in collaboration with various organizations, including the Masakhane Research Foundation network, Lelapa AI, Mozilla Common Voice, and EqualyzAI, to create a growing ecosystem of African language models and data.
  • The goal is to create a balanced, publicly available dataset of African languages, allowing for the development of useful AI tools, such as chatbots, education tools, and local service delivery platforms.

Statistics:

  • The project aims to collect 100 hours of speech data.
  • 5 languages in Kenya (Dholuo, Maasai, Kalenjin, Somali, and Kikuyu) and 5 languages in Nigeria (Bambara, Hausa, Igbo, Nigerian Pidgin, and Yoruba) are being targeted.
  • The project is collecting data in 7 South African languages (isiZulu, isiXhosa, Sesotho, Sepedi, Setswana, isiNdebele, and Tshivenda).
  • The dataset aims to reflect authentic language use in African communities and will be used to develop models for real-world platforms.

Sources:

  • The African Next Voices project, funded by the Gates Foundation and Meta
  • Masakhane Research Foundation network
  • Lelapa AI
  • Mozilla Common Voice
  • EqualyzAI
  • Gates Foundation
  • Meta
  • Data Science Nigeria
  • Maseno University
  • Utavu AI Foundation
  • University of Pretoria

Read more