The Interspeech 2026 NeckVibe Challenge: Voice Disorder Detection via Real-World Monitoring of Neck-Surface Vibration

TL;DR — The Interspeech 2026 NeckVibe Challenge released 46,400 hours of week-long ambulatory smartphone-accelerometer voice data to detect vocal hyperfunction, yielding test AUCs up to 0.93 for phonotraumatic voice disorders and 0.86 for nonphonotraumatic disorders.

Key contributions

• Released the largest real-world ambulatory voice monitoring dataset to date, comprising 46,400 hours from 582 participants across one week of daily activities.
• Benchmarked two distinct clinical binary classification tasks: phonotraumatic voice hyperfunction (PVH) and nonphonotraumatic voice hyperfunction (NPVH) detection.
• Demonstrated that preserving within-day temporal structure and utilizing frame-to-frame dynamic deltas (delta and delta-delta) and feature ratios significantly improves PVH classification.
• Revealed that NPVH detection is inherently more difficult, benefiting more from global subject-level distribution metrics and context separation (speech versus singing) than within-day variability.

Problem

Clinic-based voice evaluations capture only brief snapshots that fail to represent real-world communicative and environmental voice usage, hiding the behavioral triggers of vocal hyperfunction (VH). Prior ambulatory monitors relied on summary-only logging due to on-device memory constraints, discarding the raw dynamics of neck skin-surface vibrations. Furthermore, existing models primarily used simple linear classifiers instead of leveraging rich temporal structures, leaving a major gap in reliably detecting both phonotraumatic (PVH) and nonphonotraumatic (NPVH) forms of VH from continuous tracking.

Method

The challenge dataset was captured using a smartphone system coupled to a lightweight uniaxial accelerometer (Knowles BU-27135) mounted below the thyroid prominence, recording neck-surface vibrations at 11,025 Hz (16-bit). Data were processed into 50 ms non-overlapping frames yielding 14 voice measures across three tiers: estimated SPL (via microphone-ACC linear regression), raw waveform features (ACC magnitude, CPP, H1-H2, spectral tilt, low-to-high power ratio), and model-based impedance-based inverse filtering (IBIF) glottal airflow features (e.g., HRF, MFDR, open/speed/closing quotients).

Top-performing teams adopted contrasting architectural paradigms. For PVH detection, winning teams leveraged regularized XGBoost or logistic regression operating on windowed time-series samples or engineered dynamic sequences—calculating frame-to-frame deltas (delta and delta-delta), coefficient-of-variation-like ratio features (e.g., standard-deviation-to-mean ratios), and feature-to-feature ratios like CPP-to-H1–H2. Conversely, top NPVH entries used hybrid systems combining CatBoost on global daily/subject distribution metrics (such as Bowley's skewness, Crowley's kurtosis, Gini, and entropy) with Multiple Instance Learning (MIL) networks using attention pooling across multi-channel frame matrices.

Training recipes strictly enforced subject-level splits (80% train, 20% test across 468 vs 114 subjects) using Stratified Group K-Fold cross-validation to prevent data leakage. Inference involved day-level probability pooling averaged per subject to predict binary diagnostic status.

Experimental setup

Evaluated on 46,400 hours of ambulatory data from 582 individuals (213 PVH, 169 PVH controls, 116 NPVH, 84 NPVH controls), averaging 11.25 hours per waking day over a week. Baselines consisted of standard logistic regression models using basic frame summary statistics (mean, median, SD, IQR, 5th/95th percentiles), yielding AUCs of 0.82 for PVH and 0.78 for NPVH. Metrics reported include area under the ROC curve (AUC), accuracy, sensitivity, specificity, and F1-score.

Results

For Task 1 (PVH detection), all six competing teams surpassed the 0.82 baseline, with the top submission achieving an AUC of 0.925 (Team SR) and 0.917 (Team DD) at 84% accuracy, 79% sensitivity, and 88% specificity. Ablations demonstrated that feature-to-feature ratios (AUC 0.89) outperformed pure dynamic (0.87) and static (0.85) features, and ratio-enhanced logistic regression outperformed complex tree ensembles for certain teams.

For Task 2 (NPVH detection), four teams exceeded the 0.78 baseline, achieving a top AUC of 0.861 (Teams SM and VA). However, NPVH was substantially harder, with lower-ranked teams dropping to AUCs of 0.579 and 0.633. Unlike PVH, NPVH models did not benefit from within-day variability features, relying instead on global distributional summaries and multi-task context separation.

Limitations

The dataset suffers from a significant sex imbalance with a heavy female predominance, reflecting the clinical population but limiting immediate generalizability to male cohorts. The challenge setup provided pre-extracted time-series feature files rather than raw accelerometer waveforms, preventing exploration of end-to-end raw-signal deep representation learning. Furthermore, disease severity was mostly mild-to-moderate, which hindered separability for nonphonotraumatic cases.

Why read this

Speech and ML researchers focusing on wearable healthcare and bioacoustic modeling should read this to understand how to design robust temporal dynamics, feature ratios, and multi-instance learning architectures for week-long ambulatory sensor streams.

Code

None released (as of this page's updated date). If you are an author with a repo, please claim this entry — see CONTRIBUTING.md.

Applications

Automated remote screening, personalized diagnostic tracking, and real-time biofeedback systems for vocal hyperfunction and occupational voice disorders.

Institutions

Massachusetts General Hospital, Harvard Medical School, Universidad Tecnica Federico Santa Maria, University of Central Florida

Funding / 經費: National Institutes of Health, National Institute on Deafness and Other Communication Disorders

Related

• A Hierarchical Feature Engineering Framework for Automated Classification of Phonotraumatic and Non-Phonotraumatic Vocal Hyperfunction — same problem · relatedness 2.9/3
• Temporal Partitioning of Vocal Activity for Detecting Vocal Hyperfunction from Neck-Surface Accelerometer Data — same problem · relatedness 2.9/3
• Attention-Based Multiple Instance Learning with Tabular Stacking for Ambulatory Detection of PVH and NPVH — shared data / evaluation · relatedness 2.9/3
• Measuring Vocal Efficiency in Daily Life in Patients with Voice Disorders Using Wireless Accelerometer and Microphone Sensors — same problem · relatedness 2.2/3
• Modeling Lombard Effects in Voice Disorders Using Daily-Life Monitoring of Ambient Noise and Voice Acoustics — same problem · relatedness 2.0/3

All 950k paper pairs scored by TypeSafe Jev (scripts/related/); relatedness 0 = unrelated … 3 = directly comparable.