We are building the open speech data that Ethiopian languages do not have.
Ethiopia has more than 80 languages. Amharic alone has around 60 million speakers, more than Italian or Korean. Yet there is almost no open speech data for it.
That means no working voice assistants, no accurate transcription, no voice search, and no way for people who are not comfortable typing to use technology in their own language.
The data does not exist because nobody built the way to collect it. That is the part we solved.
Almost everyone in Ethiopia uses Telegram, so we put the whole thing inside it. No app to download, no signup form, no laptop needed.
A contributor is shown a sentence. They read it aloud and send a voice message.
Other contributors listen and vote on whether the recording matches the text.
Our pipeline measures loudness, silence, duplicates, and whether the sentence was fully read.
Contributors earn mobile airtime for accepted work, on the phone they already use.
We are not asking anyone to trust a plan. Our first batch of Amharic speech is published and free for anyone to download and check.
Every clip carries the speaker's gender, age band and region. Very few African speech corpora do. It lets a researcher measure how a model performs for older or rural speakers, instead of reporting one number and hoping.
An independent researcher, the founder of Addis AI, ran seven speech recognition systems against our held out test set and published the results. We did not ask him to.
| System | Company worth | Word error rate |
|---|---|---|
| Addis AI STT | Ethiopian startup | 25.0 |
| Google Chirp 3 | $4.3 trillion | 32.9 |
| Addis Scribe (md) | Ethiopian startup | 41.8 |
| Meta MMS-1B | $1.4 trillion | 52.0 |
Five point seven trillion dollars of market value, and the best result on Amharic came from a team in Addis Ababa. The missing ingredient was never compute. It was data.
Lower is better. Independently measured on the Dataset.ET v0.1.0 test set: 718 clips, 2.22 hours, 34 speakers held out from training. Market values as of August 2026.
We took the recordings that human validators had unanimously approved, the cleanest tier we have, and ran objective acoustic checks over them.
People cannot hear whether a recording is good enough to train a model. Machines can. That is where we need compute.
We know what the next six months look like, and compute is the only piece we do not already have.
Collected from 1,000+ contributors. 22.7 hours published as our first open release on Hugging Face.
Our national campaign is built to take collection from 33,000 recordings a month to several times that.
Every clip checked by speech recognition against the sentence it was meant to be. This is the GPU work.
Our first open Amharic model. The best measured result on our own test set is 25.0 word error rate. We intend to beat it and give the model away free.
Oromo and Tigrinya follow the same path once the Amharic pipeline is proven.
Everything else in our stack runs happily on cheap hardware. These three do not, and they are what stands between us and a corpus people can trust.
Run speech recognition over each recording and compare it to the sentence the person was asked to read. This is what catches the 24.5 percent.
Fine tune wav2vec2-BERT to build free Amharic, Oromo and Tigrinya speech recognition, retrained as the corpus grows.
Speaker diarization to find clips with a television, a radio, or a second person in the room. These quietly poison a training set.
It should be built on data that Ethiopians recorded, were paid for, and can freely use. Not scraped, not extracted, and not owned by someone far away.
We have the collection engine and the community. We need the compute to make the data good enough to matter.
dataset.et · huggingface.co/datasets/snapwre/amharic-speech