Hohe ASR scores 16.1% word error on our published Amharic benchmark. We published the benchmark first, scored six open models against it, and only then put our own number next to theirs, because a benchmark released alongside the model that wins it is marketing.
Where it is worse
On real telephone audio the same model is far weaker. Call recordings are narrowband, noisy and full of overlapping speech, and the gap between read speech and a phone call is the single largest unsolved problem in our stack. We say so on the research page rather than quoting the better number everywhere.
- Read speech, clean microphone: the 16.1% figure
- Spontaneous speech in a room: worse, and measured
- Telephone audio: much worse, and the current priority
Why publish the hard case
Because anyone evaluating us will find it in the first hour of their own testing, and the only thing a hidden weakness buys is the moment they stop believing the rest of the page.