Authors: Lakshmi Narasimhan Govindarajan¹, Sagarika Alavilli², Josh McDermott¹
¹MIT; ²Harvard University
Background: Sensory inferences are inherently uncertain because they are based on ambiguous observations. This uncertainty is likely exacerbated by hearing loss. While estimating uncertainty is critical for decision-making, the mechanisms by which the auditory system estimates uncertainty for real-world perceptual tasks—such as sound localization and pitch estimation—remain poorly understood.
Method: We developed stimulus-computable models optimized to represent probability distributions over location and fundamental frequency. The models were trained on large datasets of natural sounds, including binaural spatial renderings and excerpts of speech and music. Model uncertainty (derived from the spread of the posterior distribution) was compared to human confidence judgments. Human participants performed localization and pitch tasks and placed monetary bets (1–5 cents) to indicate their confidence in each judgment.
Results: Human confidence patterns were closely mirrored by the model across both domains. Specifically, humans placed lower bets for peripheral sound locations and narrow-band stimuli in localization tasks, as well as for complex tones containing only high-numbered harmonics in pitch tasks. These performance-dependent fluctuations in confidence were accurately predicted by the normative uncertainty estimates of the models.
Conclusion: Humans maintain internal estimates of uncertainty for sound location and pitch that are normatively appropriate. The close alignment between human confidence and optimized model predictions suggests that the brain utilizes statistical structures from natural environments to estimate perceptual reliability. This framework offers a generalizable approach for studying confidence and uncertainty across various auditory and perceptual domains.
