Listeners may be able to recognise speech in adverse condi- tions by "glimpsing" time-frequency regions where the target speech is dominant. Previous computational attempts to iden- tify such regions have been source-driven, using primitive cues. This paper describes a model-driven approach in which the like- lihood of spectro-temporal patches of a noisy mixture represent- ing speech is given by a generative model. The focus is on patch size and patch modelling. Small patches lead to a lack of dis- crimination, while large patches are more likely to contain con- tributions from other sources. A "cleanness" measure reveals that a good patch size is one which extends over a quarter of the speech frequency range and lasts for 40 ms. Gaussian mixture models are used to represent patches. A compact representa- tion based on a 2D discrete cosine transform leads to reasonable speech/background discrimination. Index Terms: speech separation, glimpsing, model-driven, spectro-temporal patches.