Abstract: Traditional neural network always adopts gradient descent method as search method, which is easily fall into local optimum. Quantum genetic algorithm was introduced into neural network, and an improved quantum genetic neural network based speech watermark algorithm was proposed in this paper. Firstly, identify a rough solution by using the improved quantum genetic algorithm to optimize the initial coefficient of BP neural network. Then search out the optimal coefficient of BP neural network with the use of gradient descent method, thus raise the accuracy of network convergence. Theoretical analysis and experimental results show, compared with traditional BP network and GA based network, our method can achieve smaller error and bigger watermark capacity.