This paper deals with a noise reduction algorithm which uses the binary masking approach in the time-frequency domain to improve speech intelligibility. In the binary masking approach, the noise-corrupted speech is decomposed into time-frequency units. Noise-dominant time-frequency units are removed by setting the corresponding binary masks as "0"s and target-dominant units are retained untouched by assigning mask "1"s. We propose a binary mask estimation by comparing the local signal-to-noise ratio (SNR) to a threshold. The local SNR is estimated by a training-based approach. An optimal threshold is proposed, which is obtained from observing the distribution of the training database. The proposed method is evaluated by normal-hearing subjects and the intelligibility scores are computed by counting the number of words correctly recognized.