# from the proj-accents repo root
mkdir -p models/cv26-hubert-svm-calibrated
hf download miquelt-9/cv26-hubert-svm-calibrated \
  --include "model.joblib" "metadata.json" \
  --local-dir models/cv26-hubert-svm-calibrated

Then start the FastAPI backend (uvicorn backend.app:app ...). On first inference, Transformers downloads the encoder BSC-LT/hubert-base-ca-2k into the local HF cache.

Pipeline

Encode audio with BSC-LT/hubert-base-ca-2k (frozen; not fine-tuned in this artifact)
Pool last hidden states with mean + std
Run this calibrated SVM → five dialect probabilities

Metrics (research snapshot)

Metric	Held-out CV26	External AINA
Accuracy	~50%	~50%
Macro F1	~51%	~50%
Top-2 accuracy	~72%	~70%

Trained on 1,440 balanced Common Voice 26 Catalan clips (96 speakers × 3 clips × 5 dialects). Suitable for a local research prototype, not a polished public release.

License and attribution

This repository / artifact (model.joblib + metadata.json): MIT. Sklearn head only; it does not redistribute HuBERT weights.
Speech encoder BSC-LT/hubert-base-ca-2k: Apache License 2.0. Copyright (c) 2025 Language Technologies Unit, Barcelona Supercomputing Center. See the upstream model card and the project NOTICE.

If you later publish fine-tuned HuBERT checkpoints derived from BSC’s weights, those weights remain under Apache-2.0 and must keep BSC attribution.