blas-m3-lm - Irish (Gaeilge) speech-to-text
Open-source Irish ASR: a wav2vec2-CTC acoustic model with a bundled KenLM 5-gram language model
(Wav2Vec2ProcessorWithLM). On BlasBench - the open Irish ASR
benchmark - this is #1 on Common Voice among all measured systems, ahead of Microsoft Azure and every
open and commercial model evaluated.
Try it on any device, no install: the demo Space runs this exact model - open the URL on your phone or laptop.
Results (BlasBench, irish normaliser)
| Test set | WER | CER | n |
|---|---|---|---|
Common Voice 25.0 ga-IE | 19.95 | 8.42 | 874 |
FLEURS ga-IE | 48.05 | 24.20 | 842 |
Numbers are read from the harness' results.json. The bundled KenLM gives -5.6 WER over the acoustic
model alone. For reference, ABAIR/Fotheidil self-report 19.6 CV with their full pipeline (different
normaliser) and 23.7 without LM - this model beats their no-LM number under a shared, controlled benchmark.
Usage
from transformers import AutoProcessor, Wav2Vec2ForCTC
import torch, librosa
proc = AutoProcessor.from_pretrained("jyoutir/blas-m3-lm") # Wav2Vec2ProcessorWithLM (KenLM bundled)
model = Wav2Vec2ForCTC.from_pretrained("jyoutir/blas-m3-lm").eval()
wav, _ = librosa.load("audio.wav", sr=16000)
logits = model(proc(wav, sampling_rate=16000, return_tensors="pt").input_values).logits
print(proc.batch_decode(logits.detach().numpy()).text[0])
The LM decoding needs pip install pyctcdecode kenlm
(on Windows: conda install -c conda-forge kenlm; Python ≤ 3.12). The KenLM travels inside the repo's
language_model/ folder, so AutoProcessor wires it up automatically.
Model
- Acoustic: wav2vec2-XLS-R (300M) fine-tuned for Irish (~1,491h gold + broadcast silver).
- LM: 5-gram KenLM built from human Irish text only (DOEGEN gold, Oireachtas
ga, gold transcripts; ~20.5k lines). Common-Voice test sentences were stripped (leakage guard).pyctcdecode,beam_width=100, α=0.5, β=1.5.
Roadmap
A fully on-device / in-browser build - distilling the LM into the acoustic weights so the same quality runs offline on a phone or laptop with zero dependencies - is in progress.
License & citation
Apache-2.0. Part of the Blas Voice / BlasBench Irish ASR project.
@misc{blasbench2026, title={BlasBench: an open Irish ASR benchmark},
author={Jyoutir Raj and John Conway}, year={2026},
howpublished={\url{https://github.com/jyoutir/blasbench}}}