ሆሄ · Hohe
Speak Amharic. Watch it written.
This is our Amharic speech recognition model running live. Tap the microphone, say something, and the words appear while the model is still working. Nothing to install.
Getting ready
Before you speak. What you record here may be used to improve Amharic AI as part of our datasets. Your name and phone number are not collected and never travel with it. Full detail in our privacy policy and terms. To have something removed, write to [email protected].
What to expect
It is good at one person speaking clearly: 16.1% word error on speakers it has never heard. It holds up down a phone line, across five regional dialects, and on the public FLEURS benchmark.
It is poor at two people talking at once, around 47% word error. It spells numbers out in words rather than digits, so twelve comes back as አስራ ሁለት. It writes no punctuation. It is Amharic only.
The speed is the part people do not expect. It reads the audio in a single pass instead of generating word by word, which makes it about five times faster than a Whisper of similar accuracy and means it does not invent sentences out of silence. The machine answering you has four ordinary processor cores and no GPU.
Take it and use it
The model is open under CC BY 4.0, with the language model, the full evaluation and usage code: huggingface.co/snapwre/hohe-asr-amharic.
It is also on Telegram, where it takes voice notes and hands back the text: t.me/dataset_ai_bot. The story of how it was built is in the write up.
If it gets something wrong, we would like to know. That is how the next one gets better, and we cannot find those cases from Addis on our own. Tell us at dataset.et/contact.