New Research Compares AI Architectures for African Language Speech Recognition

This empirical study systematically evaluates various AI architectures and adaptation strategies for Automatic Speech Recognition (ASR) in low-resource African languages. Researchers compared three distinct models: Phi-4-multimodal-instruct, Whisper-large-v3-turbo, and MMS-1B-all, across five low-resource African languages and Swahili, which served as a mid-resource benchmark. The investigation focused on understanding the impact of data sufficiency, monolingual versus multilingual adaptation, domain-specific fine-tuning, and the transferability of language adapters.

The findings indicate that Phi-4 achieved the highest overall accuracy, while MMS demonstrated comparable Character Error Rates (CER). Whisper, however, exhibited high CERs, consistent with its known tendencies for hallucinations and repetitive outputs. A critical discovery was that performance saturation occurs at varying data scales for different models, providing valuable insights into data efficiency for developing ASR systems for these languages. The study also highlighted that the typological complexity of a language consistently influences error rates.

Further analysis revealed nuanced effects of domain fine-tuning: it improved performance for Phi-4 and Whisper but unexpectedly degraded MMS, suggesting limitations in its adapter-based adaptation approach. Notably, using proxy language adapters proved more effective than training MMS adapters from scratch. This research offers practical guidelines for developers and researchers aiming to build more robust and accurate ASR systems for the diverse linguistic landscape of Africa, directly addressing the challenges posed by low-resource languages.