The invention provides a voice identifying method based on deep neural network characteristic training. The method involves the realization and identification of Gabor filter bank characteristics, Gabor filter sub-banks, and a deep neural network (DNN), which is achieved through the following steps: a Gabor filter extracting automatic voice identifying characteristics from a voice signal, firstly on the basis of a distributive voice identification standard, extracting a logarithm Me 1 spectrogram from the voice signal, then conducting convolution on the spectrogram and each 2D filter from the Gabor filter bank; selecting a specific modulation frequency, such that a transfer function of the filter exhibits constant overlapping in a modulated frequency field; an automatic voice identifying system, on the basis of character error rate in a test set, carrying on evaluation, and finally acquiring an identification result. According to the invention, the Gabor filter sub-bank can reduce character and word identification errors, and exhibit the channel distortion resistance and low signal-to-noise ratio. The method uses a voice identifier having a high time modulation filter, has low error rate, and increases the distinctiveness among object types.