SunflowerASR sets new error-rate record for African language speech recognition

Researchers released SunflowerASR, an encoder‑decoder speech recogniser that uses Group Relative Policy Optimisation (GRPO) to reduce errors. The model builds on Whisper‑large‑v3 and was tested on 51 African languages. It achieves the lowest published macro‑averaged word error rate (WER) on the AfriVox‑v2 benchmark (0.227) and on SimbaBench (0.372). The gains are strongest when training data are limited or audio is noisy.