The Data Guy

Hohe ASR is out. Open Amharic speech to text. Start with the speed, because it is the part people do not expect. On four ordinary CPU cores, a five second voice note is transcribed in 0.39 seconds. A thirty second clip takes 7.75 seconds. An hour of audio goes through in under ten minutes, on a machine that costs about nine cents an hour to rent. No GPU anywhere in that sentence. It is a CTC model, one single pass over the audio instead of generating word by word, which makes it roughly five times faster than a Whisper of similar accuracy. It also means it does not invent sentences out of silence. Accuracy: 16.1% word error and 5.4% character error on speakers it has never heard. 880 hours of training audio, including five regional dialects and a third of it pushed through a phone line on purpose. What it is bad at, plainly: two people talking at once, around 47% word error. Numbers come out as words. No punctuation. Try it now, nothing to install: @dataset_ai_bot Weights, open under CC BY 4.0: huggingface.co/snapwre/hohe-asr-amharic Send us what it gets wrong.