New AI Optimization Method Significantly Boosts Speech Recognition for Low-Resource African Languages

Researchers have developed a novel approach, Group Relative Policy Optimisation (GRPO), to enhance speech recognition models, particularly for languages with limited data. While GRPO has previously been applied to large language model-based systems, this new work extends its benefits to smaller, non-LLM encoder-decoder Automatic Speech Recognition (ASR) models.

Using Whisper-large-v3 as a foundation, the team specifically applied and tested this method across 51 African languages. The findings indicate that GRPO is especially effective in scenarios where training data is scarce or audio quality is poor, which is often the case for many African languages.

The resulting model, named SunflowerASR, has achieved state-of-the-art performance, recording the best published macro-averaged Word Error Rates (WER) on two key benchmarks for African languages: AfriVox-v2 (0.227) and SimbaBench (0.372). This open-source availability of SunflowerASR marks a significant advancement for multilingual speech technology in Africa.