Variability During Learning is Not Universally Beneficial: Effects of Task and Type of Variability

Do we learn better from a broad range of variable examples, or a narrow set of consistent ones? While higher-variability training is often assumed to generate more robust category learning, it remains unclear how different types of variability affect task performance, and whether variability affects all tasks equally. We examined how different forms of variability influence learning by manipulating pitch variation (but equating average pitch) across shorter and longer timescales when participants learned to identify new voices. To examine task-specific differences, we tested performance on a voice identification in noise task and a speech intelligibility task. The nature of variability during training produced distinct outcomes for subsequent voice recognition in noise: exposure to fine-grained, moment-to-moment pitch variability during learning affected performance more than exposure to broader, between-sentence pitch variability. In contrast, neither form of variability affected speech intelligibility: participants showed a comparable benefit to speech intelligibility when listening to trained voices compared to novel voices, regardless of their previous exposure to pitch variation within trained voices. These findings demonstrate that variability is not uniformly beneficial, but instead interacts with the type of variability and the task. These results challenge current theories of perception and learning, which either propose that variability is stripped away during learning or treat variability as a single construct. Instead, our results demonstrate that at least some types of variability are maintained during learning—but not all types of variability are treated equally, and different types are utilised in distinct ways for different tasks.