What are the properties that neural representations must have in order for them to support successful behaviour? In many fMRI studies, it has been found that “more is better”: Stronger neural activation predicts improved behavioural performance. However, for discrimination tasks, what matters is how distinguishable the neural representations are,rather than the total intensity of activation. For example, in order for the difference between the phonemes /r/ and /l/ to be perceptible, the neural patterns evoked by the two sounds must be distinct. I will describe how, by analysing the multi-voxel spatial fMRI patterns elicited by these stimuli in English and Japanese speakers, the statistical separability of such neural representations can be directly quantified. Moreover, in right primary auditory cortex, the separability of these fMRI patterns strongly predicted the degree to which subjects could behaviourally discriminate the stimuli that gave rise to them. Measures of the separability of neural representations could potentially be broadly applicable as indicators of the representational competence underlying, but independent of, behavioural performance. I will discuss possible future directions, including application to reading ability and predictions for training studies.