Context-dependent auditory inference in environmental sound recognition

In everyday listening, humans make multiple perceptual inferences from sounds (e.g., what object was that? Was it heavy? Hard? Metallic?). Such judgments are robust across different listening contexts (e.g., different motions, etc.). Using listening tasks and Ideal Observer Models trained on a large set of acoustic cues, we investigated which cues might underlie human auditory physical inference. Specifically, we asked listeners to identify attributes of objects (e.g., shape, weight, material, etc.) from clattering sounds. Prior studies of speech perception have highlighted the importance of context-dependent inference (i.e., the use of different acoustic cues in different contexts, such as different speakers, accents, etc.), and we investigated its role in such tasks. Our stimuli varied across multiple physical attributes and thus, for any query (e.g., weight) the other attributes (shape, material) served as unknown and variable contexts. Additionally, we presented stimuli in different “acoustic contexts,” manipulating spectro-temporal structure and bounce patterns to further elucidate the role of various acoustic cues in shaping human judgments. Models show different acoustic cues best predict human judgments in different conditions, suggestive of context-dependent inference. The use of such mechanisms in non-speech categorization suggests they are not specific to speech perception but may instead be fundamental to hearing.