Mamba ASR Achieves Conformer-Level Accuracy with Lower Compute in South African Languages

Mamba state space models match Conformer-based ASR accuracy across seven South African languages using 50 hours of speech per language. The Mamba architecture offers faster training and lower computational requirements, though both models show similar generalization struggles with long-form audio.