Unsupervised clustering of ASMR triggers: a comparative study of tensor decomposition and PCA

Autonomous Sensory Meridian Response (ASMR) provides psychological and physiological benefits such as stress reduction and sleep facilitation. However, the objective acoustic properties defining its diverse triggers remain underexplored, as current taxonomies rely heavily on subjective labels that often confound sound texture with temporal dynamics. This study elucidates the latent acoustic structure of ASMR content by contrasting static spectral analysis with multi-window temporal modeling to establish an objective, signal-based taxonomy. Using a large-scale dataset of approximately 3,000 ASMR video segments, we compared a static approach using Principal Component Analysis (PCA) on time-averaged acoustic features against a dynamic approach using Tucker tensor decomposition, which preserves the multidimensional (Video × Time × Feature) structure of the audio signal. The static PCA approach successfully differentiated triggers based on timbral “texture,” isolating speech, whispering, dry mechanical noises, and wet mouth sounds. Conversely, the dynamic tensor decomposition categorized triggers based on their temporal “structure.” It revealed that a vast majority of triggers—including those traditionally grouped as “touch”—are fundamentally unified by their stationary and repetitive temporal behavior, regardless of their specific sound source, a pattern further confirmed by an objective temporal variability measure and by a blinded two-annotator reliability check (Cohen's κ = 0.738 for texture, 0.536 for stationarity). A quantitative comparison of the two methods yielded a low Adjusted Rand Index (ARI = 0.0447), which, while statistically distinguishable from chance agreement (permutation test, p <0.0001), remains practically negligible in magnitude. These findings demonstrate that sound texture and temporal structure represent distinct but complementary dimensions in ASMR acoustics. We propose a dual-axis framework for ASMR clustering integrating both static and dynamic features, offering a robust foundation for future content recommendation systems and therapeutic applications.