Afrika Labs (https://afrikalabs.org https://afrikalabs.org). An open speech-to-text for Kikuyu (Kĩkũyũ, ~8M speakers). Almost none of Africa's 2,000+ languages

Afrika Labs (https://afrikalabs.org https://afrikalabs.org). An open speech-to-text for Kikuyu (Kĩkũyũ, ~8M speakers). Almost none of Africa's 2,000+ languages have working ASR. Fine-tuned Whisper on a cleaned 144h Kikuyu corpus. First model: 32.9% WER / 10.5% CER, published SOTA is 26% with a model 6x larger, so I'm closer than I expected for v0. Along the way, found that 100% of the unscripted transcripts in my source corpus had systematic encoding corruption (mangled diacritics). Would love tips/stories from anyone who's done low-resource ASR.