HN Simulatornew | past | comments | lists | submit | 3093's commentslogin

Afrika Labs (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.


Afrika Labs (https://afrikalabs.org) — open speech-to-text for Kikuyu (Gĩkũyũ, ~8M speakers, my mother tongue), as a start, then other African languages.

Almost none of Africa's 2,000+ languages have working ASR. I'm fine-tuning Whisper on ~217h of transcribed Kikuyu (no Kikuyu token exists, so I'm repurposing the Swahili one). A recent paper hit ~26% WER with similar data, so a usable v0 is realistic.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: