Grant title: Natural Sound Statistics in Auditory Scene Analysis
Principal (co-)investigators: Jan Schnupp, Josh McDermott (MIT), Nicol Harper (Oxford)
Grant scheme: General Research Fund grant Nr 11100617
Funding amount: HK$ 928,200
Grant duration: 01.01.2018 - 31.12.2020
Abstract: Picture yourself on a sunny weekend afternoon, sitting by the edge of a stream, water rushing over a small weir, geese honking in the background, and a bumble bee buzzing past. All of these sounds are a constant presence, but drift in and out of your consciousness only fleetingly, and none prevent you from enjoying a conversation with friends. Afternoons like those enrich our lives, we accept them gratefully, and few of us wonder how it is possible that we can so effortlessly shift our attention from the conversation to the buzzing bee or the rushing water or the honking geese, given that all these sounds arrive at our eardrum as an impenetrable mish-mash. So how does normal, healthy hearing take in these sounds all separately and yet all at once? We believe that neurons in the brain may be sensitive to key, identifying statistical properties of the various sound sources, and that this will make them respond selectively to only those sound sources that match the statistics they are tuned to. We also believe that two of the most important statistical properties are "sparseness" and "correlation structure". The geese make occasional but relatively loud honks, interspersed with periods of silence. This we would describe as "sparse". Rushing water sound is less sparse because its intensity varies much less over time. If neurons prefer sparse sounds, then these would respond strongly to the sound of geese but not the water, and their responses would thus to an extent separate the sound of the geese from the rest of the auditory scene. To distinguish the gurgling water and the buzzing bee, but those you could separate using "frequency correlations". Water sounds involve countless bubbles gurgling. Large bubbles make deeper sounds, small bubbles make higher sounds, and high and low bubble sounds are not synchronized. In contrast, each wing beat of a bumble bee is a "flap", in which many frequencies all go off together, perfectly correlated. Neurons which require high frequency correlation would hear mostly the bee, neurons requiring weak correlation hear mostly the water. We hypothesize that sensitivity to these key statistical properties emerges gradually along the ascending auditory pathway, and we propose to investigate neural responses from key stages of the ascending pathway, including the inferior colliculus, primary and secondary auditory cortex of rats to synthetic auditory texture sounds with systematically varying statistics, presented in isolation and superposition.