One of the strengths of this dataset is that it includes various sounds produced by the same sources (within-class acoustic variability) as well as acoustically similar sounds produced by different sources (between-category confusion), allowing for disentangling category-level semantic processing from acoustic similarity effects. Using our dataset, researchers can selectively sample near-boundary examples across categories and investigate the mechanisms driving formation of category decision boundaries and misclassifications.
By providing acoustically varied examples within each category, the dataset also supports development of more robust classifiers that capture invariant categorical features rather than stimulus-specific properties. Our dataset's within-category variability seems particularly valuable for implementing and evaluating domain adaptation techniques (Meza, Habets, Muller, & Sarti, 2020), or data augmentation strategies (Salamon & Bello, 2016). Exemplar variability is critical for improving model generalization and for benchmarking machine learning systems aimed at naturalistic sound recognition. Furthermore, our dataset’s rich structure, along with human identifications and behavioural ratings could be used as another dimension enriching model recognition performance. A similar approach incorporating semantic descriptions into deep neural network training improved model performance and provided better approximation of human behaviour than other models (Esposito et al., 2024).
Clinically, our dataset holds promise for enhancing diagnostic and rehabilitative protocols for auditory disorders. For instance, our dataset's acoustic diversity within categories enables more nuanced assessment of cochlear implant (CI) users' perceptual capabilities. This would be a major improvement upon current practices which typically involve testing audibility of tones or speech and offer no insight into perception of individual environmental sounds or complex soundscapes (Lelic, Picou, Shafiro, & Lorenzi, 2025). Similarly, individuals with auditory agnosia, who exhibit deficits in recognizing sounds despite preserved hearing, could benefit from diagnostic tests that incorporate acoustically diverse exemplars and various sounds produced by the same source.
Please get in touch if you would like to add sounds, features or have new ideas on how to expand or use the dataset.
Last modified 2025/09/17
Designed by @mkachlicka