CHI-TUM/CoughPhase-CLR

<div align="center"> <h1>CoughPhase-CLR</h1> <p><em>An acoustics-informed foundation model for coughing sound classification</em></p> </div>

CoughPhase-CLR is a self-supervised foundation model for coughing sound classification. Unlike generic contrastive frameworks that build positive pairs through random cropping, it constructs positive pairs from the distinct physiological phases of a single cough (explosive phase vs. intermediate + voiced phases). Built on OPERA-CE (an EfficientNet-B0 encoder trained with a contrastive objective), it is pretrained on ~40 hours of public cough audio and evaluated on five downstream tasks spanning COVID-19 detection, COPD-state classification, gender, and smoker-status prediction. Phase-aware pretraining consistently outperforms standard random cropping on coughs and is more data-efficient. <div align="center"> <img width=80% src="https://cdn-uploads.huggingface.co/production/uploads/68643b761bd37a5e3db83c0f/PxzBkJxuW6JXQbZXiy2aw.png"> </div>

The code is available at: https://github.com/your-repo/CoughPhase-CLR

Checkpoints

Two EfficientNet-B0 encoders, same OPERA-CE architecture and same ~40 h of cough pretraining data — they differ only in how contrastive positive pairs are built:

`CoughPhase-CLR`: The proposed model, pairing the two physiological phases of a single cough (explosive vs. intermediate + voiced)

`OPERA-CE-Cough`: The baseline, using OPERA-CE's original random-crop pairing

Citation

Kindly cite our work if you find it useful.

@misc{moldovan2025coughphaseclr,
      title={CoughPhase-CLR: Designing an acoustics-informed foundation model for coughing sound classification},
      author={Marius Moldovan and Anton Batliner and Thomas M. Berghaus and Bj\"orn W. Schuller and Andreas Triantafyllopoulos},
      year={2026},
}