Blas-IE: the #1 open Irish speech-to-text model
Our goal was simple:
- Build the best Irish speech-to-text model in the world.
- Open-source it - release the model weights publicly, so anyone, anywhere can download and run it for free.
We did both. Blas-IE is #1 on BlasBench, ahead of Microsoft Azure, ElevenLabs, and every open model we tested. This post goes over how we did it.
19.95% WER - #1 on the BlasBench Common Voice leaderboard.
The goal
From benchmarking the existing models on BlasBench, the ones that worked best were closed source - locked behind paywalls by big tech companies. The open ones were far behind on Irish: OpenAI's Whisper, for example, is useless for it. We wanted to change that, with one model that is genuinely state of the art and fully open - so the best Irish speech tech belongs to everyone.
This is our Irish speech-to-text model: it turns spoken Irish into text. More model releases are lined up as we continue our research.
What we built
Blas-IE beats every system we measured - open or commercial. The full table, FLEURS results and confidence intervals live on BlasBench - our open benchmark, where anyone can test our model or score their own against the exact same bar.
How we built it
Blas-IE starts from wav2vec2-XLS-R, a model that has already been pretrained on a huge amount of speech - in our testing, the strongest base to build on. It has learned the shape of human speech across dozens of languages (but not Irish!). So our job was to teach it Irish.
To do that we needed data - and since there is very little transcribed Irish in the world, we had to build our own. We took great care to select a small core of open, human-labelled Irish, then gathered hundreds more hours of Irish audio from radio, TV and podcasts and labelled it using a proprietary, pseudo-labelling-adjacent technique - reaching around 1,500 hours of training audio. We fine-tuned the model on all of it.
Once we had the data, the training itself was cheap. Blas-IE trained on a single A100 on RunPod - about 40 hours, for roughly $110 of credits. State of the art Irish speech recognition for a small price, not the thousands a big lab would spend.
The biggest single jump, though, didn't come from training at all. Using an automated research loop - it proposes small experiments, scores each one on BlasBench, and keeps whatever wins - we turned up cheap tricks like this one: a small language model bolted on at decode time that nudges the output toward real Irish words.
A single decode-time trick cut WER from 25.59% to 19.95% - that took us past Microsoft Azure and into first place.
Conclusion
We open sourced the best Irish speech-to-text model in the world, free for anyone to use. Download it on Hugging Face at jyoutir/blas-m3-lm - you can run it anywhere.
Try it yourself at blasapp.com/research/dictate.
What's next
Our next research effort is to create the first ever Irish speech to English text model, and open source that too - while we keep improving our models and upholding our values of open, free research.
