Core dimensions of the perceptually relevant voice identity space

Human listeners are remarkably skilled at identifying voices, yet the mental representations that support voice identity processing remain poorly understood. In this study, we present a data-driven approach that reveals core dimensions of the voice identity space. We compiled a large-scale voice corpus including 337 male speakers producing each a unique sentence (VocDat). Using a Sparse Positive Similarity Embedding (SPoSE) model trained on voice similarity judgments from 389 participants, we identified a five-dimensional voice identity space capturing human between-speaker similarity judgments. Regression analyses showed that these model dimensions were characterised by distinct combinations of acoustic voice properties, with pitch, spectral slope, and syllable rate emerging as the most consistent and informative features. These acoustic labels closely aligned with perceptual voice attributes provided by trained phoneticians (e.g. low pitch, old age, fast speech), demonstrating that these dimensions can be interpreted in a perceptually meaningful way. By combining a data-driven modelling approach with a large and acoustically diverse speaker sample, and by characterising voice identity space dimensions in both acoustic and perceptual terms, our findings provide perceptually grounded insights into the structure of voice identity representations.