How Architecture, Typology, and Adaptation Strategies Shape ASR for Low-Resource African Languages: A Controlled Empirical Study

A controlled study comparing Phi-4-multimodal-instruct, Whisper-large-v3-turbo, and MMS-1B-all across five low-resource African languages shows Phi-4 achieving superior accuracy. Results indicate that domain fine-tuning improves Phi-4 and Whisper performance but degrades MMS, while typological complexity remains a primary driver of character error rates.