Monaural speech separation has been adopted in other applications, e.g. paralinguistics, hearing aids, online video conferences, human–machine interactions, and speech recognition. In recent years, deep learning has replaced previous methods, e.g. Gaussian Mixture Models, the hidden Markov, the independent component analysis, and the non-negative matrix factorization for utterance separation tasks. The time-frequency separation speech using masking as a training goal offers state-of-the-art performance. In this paper, we propose a multi-mask learning and vector training method. First, the network is adopted to teach the deep embedding vectors. These embedding vectors are adopted as the input features for another backbone network with the multi-mask target. The knowledge is accumulated by learning many time-frequency masks and deep embedding vectors. Experimental results have shown that our multi-mask learning and vector training model achieves higher performance than the training approach with a single mask and the multi-mask learning approach without embedding feature vectors.