A new approach to the design of time windows is presented for detection of transient-evoked otoacoustic emissions (TEOAE). The windows are designed with reference to a minimum mean square error criterion involving the correlation properties of the ensemble of responses. Latency information is introduced in the detection process by windowing at different scales that result from wavelet decomposition. The significance of both subject- and population-specific time windows is investigated. The detection performance is evaluated on a health screen database consisting of 4989 records. The results show that the present Approach to windowing yields a significantly better performance in separating normal-hearing subjects from hearing-impaired subjects when compared to detection based on unwindowed signals. With time windowing, the specificity increased with almost 15% at a fixed sensitivity of 90%.