Notes on Somali speech recognition, the Afar language gap, dialect and ASR, and language technology across the Horn of Africa.
We tested the leading commercial speech recognition systems on Somali audio. The best one gets nearly six out of ten words wrong. Here is why.
Afar is spoken by millions across Djibouti, Ethiopia, and Eritrea. There is no speech recognition dataset, no model, and no commercial coverage for it anywhere.
A speech recognition model trained on standard Somali does not reliably transcribe Djiboutian Somali. Here is why dialect, not language, is the real unit of the problem.
How we built the first ground-truth word error rate benchmark for Somali speech recognition, what we found, and what the results do and do not show.
Tens of millions of people in the Horn of Africa speak languages with little or no coverage in modern language technology. This is what that gap looks like in practice.