Afrika Labs
An open speech-to-text model for Kikuyu (~8M speakers), built by fine-tuning Whisper on a cleaned 144-hour Kikuyu corpus. The first model achieves 32.9% WER / 10.5% CER, close to the published SOTA of 26% from a model 6x larger. Part of a broader effort to bring working ASR to Africa's 2,000+ languages, most of which lack it; the author also found systematic encoding corruption (mangled diacritics) in the source corpus's unscripted transcripts.
What it addresses
Almost none of Africa's 2,000+ languages have working ASR; Kikuyu lacked an open speech-to-text model.
Shared over time
Each entry reflects what was shared on that date. Details may have changed since then.
Afrika Labs
An open speech-to-text model for Kikuyu (~8M speakers), built by fine-tuning Whisper on a cleaned 144-hour Kikuyu corpus. The first model achieves 32.9% WER / 10.5% CER, close to the published SOTA of 26% from a model 6x larger. Part of a broader effort to bring working ASR to Africa's 2,000+ languages, most of which lack it; the author also found systematic encoding corruption (mangled diacritics) in the source corpus's unscripted transcripts.
Problem: Almost none of Africa's 2,000+ languages have working ASR; Kikuyu lacked an open speech-to-text model.
Posted by @3093 · Original comment ↗
“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.”