WAXAL-NET: Fine-Tuned Edge ASR Across 19 African Languages

Multilingual speech foundation models such as Whisper, MMS, and Omnilingual ASR claim recognition coverage spanning hundreds to thousands of languages, yet this coverage remains largely unrealized for African languages, constrained by a parity gap in recognition quality and an efficiency gap that makes billion-parameter models impractical for edge deployment. We evaluate whether compact, domain-specialized models can close both gaps using WAXAL, a 2,279-hour corpus of spontaneous, image-prompted speech across 19 African languages. Fine-tuned edge models (39M-300M parameters) achieve a macro-averaged WER of 38.0% compared to 64.9% for the best zero-shot foundation baseline, a 26.9-point reduction using models 3-40x smaller. Cross-domain evaluation on FLEURS shows fine-tuned models recover usable performance on out-of-distribution speech, while zero-shot models regain their advantage only when the test domain matches their pretraining distribution, indicating that domain match, not scale, is the primary driver of performance. A distributed native-speaker audit across all 19 languages reveals systematic architectural failure patterns: autoregressive Whisper models are prone to unbounded repetition loops, while CTC-based MMS-300M fails through recoverable phonetic approximation instead. We further show that WER alone misrepresents performance for syllabary-script languages, where CER/WER ratios reveal substantially higher character-level accuracy than headline WER suggests. We release all model weights, fine-tuning and evaluation scripts, and a cleaned WAXAL subset covering all 19 languages to support future African ASR research.