This research work proposes a new Genetic Algorithm (GA) based fusion scheme to effectively fuse the Time–Frequency (T–F) binary mask of voiced and unvoiced speech. The perceptual cues such as correlogram, cross-correlogram and pitch are commonly used to obtain the T–F binary mask of voiced speech. Recently, researchers use speech onset and offset to segment the unvoiced speech from the noisy speech mixture. Most of the research work which uses speech onset and offset to represent the unvoiced speech, combine the segments of unvoiced speech with the segments of voiced speech to obtain the T–F binary mask. This research work effectively fuses the T–F binary mask of voiced and unvoiced speech, instead of combining the segments of voiced and unvoiced speech using a Genetic Algorithm (GA). Moreover, a new method is proposed in this research work to obtain a T–F binary mask from the segments of unvoiced speech. The performance of the proposed GA based fusion scheme is evaluated using measures such as quality and intelligibility. The experimental results show that the proposed system enhances the speech quality by increasing the SNR with an average value of 10.74 dB and decreases the noise residue with an average value of 26.15% when compared with noisy speech mixture and enhances the speech intelligibility by increasing the CSII, NCM and STOI with an average value of 0.22, 0.20 and 0.17 as compared with the conventional speech segregation systems. © 2019, Springer-Verlag GmbH Germany, part of Springer Nature.