Extraction of a specific speech signal from convolutive mixtures of multiple speeches is a challenge since different speeches may share similar characteristics. Based on our semiblind negentropy maximization algorithm for separating multiple speech signals, we further present an algorithm for extracting a desired speech by constructing a corresponding reference signal. Specifically, two kinds of reference signals are explored, which include a clear speech from the specific speaker and a rough estimation of blind source separation, respectively. Extensive experiments with synthetic data and recorded speeches are carried out to test the performance. The results show that the proposed algorithm can nicely extract an expected speech signal but discard the other speeches.