Speech is one of the most natural ways people communicate, yet AI systems still struggle to understand and generate speech across the linguistic diversity of Africa. Differences in language, accent, pronunciation, speaking style, and available training data create significant challenges for building reliable speech technologies. The Speech Intelligence team researches the models, methods, and systems needed to make speech AI work across African languages, from automatic speech recognition and speech synthesis to voice cloning, dubbing, and spoken-language interaction.
Research teams: Speech Intelligence · Language Intelligence · Language Resources · Model Intelligence · Responsible AI
The team explores how AI systems recognise and interpret spoken language from audio signals. Its work spans automatic speech recognition, acoustic modelling, and pronunciation variation, addressing the challenges of understanding speech across African languages, accents, and speaking conditions.
Advancing natural, intelligible, and expressive speech in African languages requires models that capture more than words alone. This research area focuses on text-to-speech, acoustic speech generation, pronunciation and prosody modelling, and multilingual speech synthesis.
Many African languages remain underserved by speech technologies. By studying shared speech representations, cross-lingual transfer, and language adaptation, the team seeks to understand how models can learn across languages and apply that knowledge to languages with limited training data.
Speech technologies must work beyond controlled recording environments. This area examines acoustic robustness, speaker variation, and adaptation to changing conditions, with the goal of making speech recognition and generation more reliable across diverse speakers, accents, and real-world settings.