Manipulating Causal Uncertainty in Sound Objects.

Causal uncertainty – how sure we are in what produced a sound that we are listening to – is a fundamental aspect of auditory cognition. It is known to be a driver of affect perception, attention, and memory, among other processes. Here, we present an optimization pipeline that systematically manipulates a sound object’s intrinsic causal uncertainty by applying a set of acoustic transforms, such as scaling a sound’s pitch, amplitude, playback speed, etc. The optimization estimator attempts to produce parameter values for these transforms that modify a sound’s causal uncertainty (Hcu), as measured by the prediction confidence of an audio classification neural network, while minimizing changes to the resulting prediction labels and transform magnitudes. We then conduct a listening test with N=20 participants to confirm that the causal uncertainty changes resulting from our proposed procedure align with human perception. Though a simple approach, this work demonstrates a first step towards generative audio systems that operate along cognitive dimensions, with powerful implications for user experience design.