Baseline Performance

Baseline Performance

Models trained on the VHM ambulatory dataset have already shown strong potential for detecting vocal hyperfunction. Past studies demonstrate that accelerometer-based features collected in real-world conditions can reliably distinguish both phonotraumatic (PVH) and nonphonotraumatic (NPVH) voice disorders from matched healthy controls.

These benchmark results provide a useful reference — the challenge invites teams to improve upon them with advanced yet interpretable models

Key Baseline Results 

PVH Detection 

• Logistic regression and support vector machine (SVM) using airflow- and spectral-based features:

  • AUC 0.82 (Cortés et al., 2018; Van Stan et al., 2020)

• Nested cross-validation logistic regression using the difference between first and second harmonic magnitudes (H1-H2 kurtosis & skewness)

  • AUC 0.81 (Ghasemzadeh et al., 2024)

NPVH Detection

• Quadratic Discriminant Analysis using cepstral peak prominence (CPP) and H1-H2 (mode)

  • AUC 0.78 (Van Stan et al., 2021)

Helpful Insights

• • Early models applied linear classifiers and demonstrated feasibility (AUC 0.70–0.74) using limited features and small datasets

  • Introduction of airflow-estimation features using IBIF markedly improved discrimination between PVH and controls

  • Parameters such as spectral tilt, H1–H2 variability, and loudness distribution consistently show strong effect sizes in ambulatory monitoring

  • Models built on daily distributional statistics perform well, but frame-level or deep learning approaches may uncover additional patterns

  • For NPVH detection, CPP (voice quality) and reduced variability in harmonics are particularly effective