Non-Contact Hand Tremor Detection via Water Pouring Sound Recognition

Early detection of hand tremors is essential for diagnosing and monitoring motor disorders such as Parkinson’s disease and essential tremor. Traditional methods often rely on cameras or wearable sensors, which can be intrusive, costly, or impractical in real-world settings. This study proposes a novel, non-contact approach for detecting hand tremors using only the acoustic signatures generated during water pouring. We hypothesize that tremor-related variations, such as fluctuations in amplitude, frequency, and rhythm, are embedded in the pouring sound and can serve as reliable indicators of hand steadiness. To validate this idea, we collected a diverse dataset of water pouring sounds representing both steady and shaky hand conditions. Recordings were made using containers of different materials and sizes, across multiple participants and environments. The audio was segmented into 5-second clips and converted into log-Mel spectrograms for input to deep learning models. We evaluated four architectures (i.e., Audio Spectrogram Transformer (AST), VGGish, ResNet50, and DenseNet121) using four cross-validation strategies, including Leave-One-Container-Out (LOCO), Leave-One-Participant-Out (LOPO), Leave-One-Section-Out (LOSO), and Leave-One-Material-Out (LOMO). AST consistently outperformed the others, achieving above 99% accuracy across all validation strategies. These results highlight the promise of sound-based tremor detection for passive and unobtrusive health monitoring in daily or clinical settings.