A glimpsing model of speech perception
Martin P. Cooke · 2003
Do listeners utilise glimpses of a clean speech target in noisy backgrounds? This paper presents a model of speech perception motivated by the glimpses notion. Using a corpus of VCV sequences, listeners' identification performance is measured in a number of noise conditions chosen to vary both in global SNR and in the number and distribution of glimpses they afford. Subjects’ performance is compared to that of a computational model which employs missing data techniques to handle glimpses. A close match to subjects’ performance can be obtained by assuming that listeners require available glimpses to occupy a certain minimum extent in time and frequency.