How speech models differ on low resource African languages
How Architecture, Typology, and Adaptation Strategies Shape ASR for Low-Resource African Languages: A Controlled Empirical Study
Computation and LanguageSound
Summary
Creating speech recognition systems for African languages with limited data is hard. The authors compared three different speech AI models on five low-resource African languages and one medium-resource language. They found that some models handle data and language differences better than others, and that the amount and type of data impact accuracy. They also learned that adjusting models to match specific topics helps some systems but hurts others. Finally, languages’ unique characteristics affect how often models make mistakes.
What this means in practice
- •For speech technology developers: Improve speech systems for African languages by choosing models and data adaptation methods proven effective for these languages.
- •For language technology product teams: Design speech recognition products that target low-resource African languages by applying data and adaptation strategies shown to enhance accuracy.$Commercial implications: Enables commercial speech recognition services for previously underserved African languages with improved accuracy and efficiency.
Authors
Mercy Muchai, Kevin Chege, Nick Mumero Mwangi, Samuel Chege Maina, Felermino Ali, Elizabeth Ankrah, Ignatius Ezeani
Abstract
Despite recent advancements in ASR, identifying effective architectures, fine-tuning approaches, and data adaptation strategies for low-resource languages remains challenging. We present a systematic comparison across five low-resource African languages and Swahili as a mid-resource contrast, fine-tuning three models from distinct architectures: Phi-4-multimodal-instruct, Whisper-large-v3-turbo, and MMS-1B-all. We analyze data sufficiency, monolingual versus multilingual adaptation, domain fine-tuning, and language-adapter transfer. Phi-4 achieves the best overall accuracy, MMS shows comparable CER, while Whisper exhibits high CER, consistent with known hallucinations and repetition loops. Performance saturates at different data scales, offering practical data-efficiency targets, while typological complexity consistently shapes error rates. Domain fine-tuning benefits Phi-4 and Whisper but degrades MMS, revealing limitations of adapter-based adaptation. Proxy language adapters outperform training MMS adapters from scratch.