Training spiking neural networks (SNN) using backpropagation methods makes it possible to obtain models comparable in quality to conventional neural networks. However, this process requires significant computational resources due to the need to process input data in the form of sparse spike sequences, which increases the amount of computation and training time. In this paper, a combined method is proposed to improve efficiency. It includes three stages: preliminary training of a formal neural network with a step activation function, transferring of weights and thresholds to an SNN with a similar architecture, and subsequent additional training with the selection of a spike processing configuration. This approach allows achieving classification accuracy on audio classification task comparable to training an SNN from scratch (0.94 vs 0.96), while significantly reducing training time (by 11 times).