Automatic multilingual speech recognition for languages of sub-Saharan Africa

Starting: 09/12/2025
PhD Student: Lu Zuo
Advisor(s): Marie Tahon (LIUM)
Co-advisor(s): Élodie Gauthier (Orange), Lina ROJAS (Orange), Aghilas Sini (LIUM),
Funding: Cifre - ANRT

The aim of this thesis is to study and develop solutions for speech understanding using weakly supervised or unsupervised approaches, in order to effectively exploit available data and make systems robust to acoustic and linguistic variability.It also involves distributing open-access corpora and models so that the scientific community and local stakeholders can reuse them, thereby promoting collaboration, innovation and adaptation to specific contexts. The ultimate goal is to advance these technologies in an ethical, inclusive and sustainable manner.

The financial and usage issues associated with the development of technologies for African languages are numerous and decisive for their adoption by users. In particular, it is essential to make these initiatives more accessible and sustainable in contexts where resources are limited. Reducing development costs is a major challenge, which requires a focus on frugality in model learning. Furthermore, the opening of new markets for voice technologies in Africa represents a significant economic opportunity, enabling the use of these solutions to be extended to a wide audience.

Finally, these innovations have a considerable social impact: they promote digital inclusion, facilitate access to information and contribute to the promotion of local languages and cultures, thereby strengthening their vitality and transmission.