Abstract: Recently, using a dictionary learning method to sparse represent speech signal becomes a popular research subject. For selecting the initial dictionary size problem, it is proposed a new dictionary training method that based on new BDS model. This method established model for dictionary size according to the relationship between dictionary size and sparse ratio. This model can self-adaptively choose properly initial dictionary for speech signal. More over, it overcomes the defects of selecting initial dictionary size in K-SVD method on experience. Appling this training dictionary method added the BDS model to speech library of lab of audio and video of Taiyuan University of Technology by simulating experiments. The results are analyzed. Simulation results show that the based on BDS model speech signal dictionary construction method not only realized that choose best dictionary size adaptively, but also improved the efficiency and stability of the dictionary training.