Introduction

Aragonese speech recognition just got a serious open fine-tune: AAL-Academia/omniasr-llm-3b-aragones-v1 (created 2026-11-07T22:17:57Z), published under the Academia Aragonesa de la Lengua org. It is a supervised fine-tune of Meta’s omniASR_LLM_3B_v2 — the standard LLM variant (not Unlimited) — with a live demo at spaces/Juanpabl/asr-aragones-demo.

The “3B” label names the family; the full encoder–decoder system sits around ~4.38B parameters and runs on fairseq2 + omnilingual-asr.

What Shipped

  • Vendor / product: omniasr-llm-3b-aragones-v1 — Academia Aragonesa de la Lengua / Juan Pablo Martínez Cortés
  • Modality: open-weight automatic speech recognition for Aragonese
  • Open vs closed: model + distribution code Apache-2.0 (upstream notices retained); source corpora keep their own terms and are not redistributed
  • Runnable today: Meta Omnilingual ASR stack (facebookresearch/omnilingual-asr); demo Space for a quick listen/test
  • Concrete fact: fine-tuned on Common Voice Aragonese Scripted Speech 25.0 (v3 split: train 12.735 h / 10,264 clips); selected checkpoint step 4,750 by development WER

Data and Training

  • Train: 10,264 clips · 12.735 h · 30 speaker IDs
  • Dev: 781 clips · 0.997 h · 6 speaker IDs
  • Test: 2,232 clips · 3.010 h · 15 speaker IDs
  • Speaker IDs are disjoint across splits; paths and identical MP3s were checked for overlap
  • Fine-tune: two A100 GPUs, LR 1e-5, effective batch 32; convolutional feature extractor frozen, remainder fine-tuned
  • Checkpoint chosen before the external spontaneous-speech gold set was available

Pro Tip

Treat the spontaneous gold numbers as an early stress test, not a production SLA — the card itself flags the set as small, difficult (older speakers / dialects), and not fully audited for alignment.

How to Run It

Upstream inference follows the Omnilingual ASR project. Audio is converted to 16 kHz mono; long recordings were processed as contiguous ~17–23 s segments near low-energy boundaries. Streaming is not offered yet.

On a single 24 GB RTX 3090, batch throughput (batches of four segments) transcribed 2 h 47 min of audio in about 13 min, excluding model load — that is offline batch speed, not live conversation latency. CPU and quantized variants are not yet benchmarked on this card.

Try the Space first if you just want to hear Aragonese in / text out before standing up fairseq2 locally.

Quality Snapshot and Limits

  • Common Voice v3 internal test (2,232 clips): WER 4.34%, CER 1.11%
  • Unseen-sentence subset (1,417 clips absent from train+dev): WER 5.50%, CER 1.39%
  • Initial spontaneous gold (7 recordings, 12 min 27 s): WER ~15.87%, CER 8.09%
  • Sensitivity without the longer interview (6 recordings): WER 11.83%, CER 4.34%

Scoring normalizes case and ordinary punctuation; accents, apostrophes, hyphens, and repetitions are kept. Automatic transcripts still need human review and do not certify spelling or variety. Known risks: omitted passages, unsupported text, altered dialectal forms.

Where It Fits

  • Minority-language ASR work that starts from Meta’s multilingual omniASR LLM line instead of training from scratch
  • Distinct from generic multilingual ASR cards: this is an Academia-backed Aragonese FT with documented CV splits and a public demo
  • Research / tooling release under Apache-2.0 — keep human review in any archival or broadcast pipeline
  • Credits: Meta Omnilingual ASR team (base); Juan Pablo Martínez Cortés (Aragonese fine-tune, eval, distribution)

Conclusion

If you care about Aragonese speech tech and wanted a zero-day open FT with honest metrics, omniasr-llm-3b-aragones-v1 is the drop to clone. Sub-5% WER on the Common Voice test set, Apache-2.0 weights, and a Space to try before you wire fairseq2 — that is how minority-language ASR should ship.

—Anabel ♡