Parakeet-TDT 0.6B HD Articulation

Official checkpoint for the paper "Huntington Disease Automatic Speech Recognition with Biomarker Supervision."

Model description

This model is an articulation-aware variant of Parakeet-TDT 0.6B HD, tuned for automatic speech recognition on speech affected by Huntington disease (HD). It extends the HD-adapted base model with auxiliary supervision from articulatory biomarker labels.

What this model does

The model transcribes English read / controlled clinical speech from speakers with Huntington disease and healthy controls. It is intended as a research model for studying robust ASR under hyperkinetic motor-speech disruption and for analyzing the effect of articulatory supervision on transcription behavior.

Training

The model was initialized from charleslwang/parakeet-tdt-0.6b-HD and further adapted using parameter-efficient encoder-side adapters with an auxiliary objective based on articulatory biomarker labels, while keeping the pretrained backbone frozen.

Evaluation

On the reported HD test set, this model achieved:

WER: 6.44
Substitutions: 1.94
Deletions: 3.21
Insertions: 1.29

Intended use

This model is intended for:

research on pathological / atypical speech recognition,
benchmarking ASR on Huntington disease speech,
studying how articulatory auxiliary supervision reshapes error behavior.

It is not intended for clinical diagnosis, treatment decisions, or standalone medical use.

Limitations

Trained and evaluated on a relatively small, high-fidelity clinical corpus.
Primarily reflects controlled / read speech rather than spontaneous conversational speech.
Did not outperform the plain HD-adapted model on overall WER.
May not generalize to severe out-of-distribution impairment, other languages, or other recording conditions.

Citation

If you use this model, please cite:

@article{wang2026huntington,
  title={Huntington Disease Automatic Speech Recognition with Biomarker Supervision},
  author={Wang, Charles L. and Chen, Cady and Gong, Ziwei and Hirschberg, Julia},
  journal={arXiv preprint arXiv:2603.11168},
  year={2026},
  url={https://arxiv.org/abs/2603.11168}
}