The paper presents a method for creating a spiking neural network model with restrictions on the bit capacity of storing weight coefficients and the frequency of signal processing. It is relevant for prospective deployment on neuromorphic computing devices. The method proposed involves surrogate gradient for tuning the weights, and quantization-aware training for imposing the restrictions. Encoding audio recordings for processing by the spiking neural network is implemented so as to allow stream data processing in real time. It is based on applying fast Fourier transform, and representing the resulting coefficients by sequences of spikes with a given frequency determined by the clock frequency of the neuromorphic device. Experiments on the Heidelberg Digits dataset showed the possibility to quantize the weights and thresholds of a feedforward spike network to int8 without significant loss of accuracy: from 0.92 to 0.91 Acc.