In single-channel speech enhancement, it is essential to determine noise reduction factors to successfully remove noise while minimizing speech distortion. These factors are typically set by a function of noise power spectral density (PSD) in time-frequency domain, and the state-of-the-art algorithm also introduces additional processes to estimate speech presence probability (SPP) to further enhance the estimation. Due to many tuning parameters, however, it is not easy to implement an algorithm that reliably estimates SPP in noise varying environment. We proposed a combination of deep learning network and an effective training method to enhance the performance of the SPP estimation module. The proposed approach is regarded as a hybrid approach, with the noise reduction factor still estimated by the conventional statistic-based single channel enhancement algorithms. The advantages and disadvantages of the proposed approach compared to deep learning approach of single channel speech enhancement are also investigated.