Description
Despite rapid and meaningful advances in speech recognition, building robust ASR systems for low-resource languages remains challenging. Key questions persist: how much data should be collected and in what forms, which model architectures offer the best opportunities for training and deployment, and what adaptation strategies are most effective for new domains and languages?
In this talk, Muchai will present an architecture-aware evaluation of 53 speech models and their accuracy-efficiency trade-offs across 39 African languages. Muchai will also discuss research findings from a systematic study fine-tuning three models from distinct architectural paradigms, revealing how model architecture, language typology, and adaptation strategies shape speech recognition performance and real-world deployment outcomes. This talk draws from research on the Paza Speech Playbook and recent publications.
Learn more about Microsoft Research Africa Seminars: https://www.microsoft.com/en-us/research/lab/microsoft-research-lab-africa-nairobi/seminars/