A method for extracting sparse and energy-efficient feature sets based on Delta-Gated Spike Encoding (DGSE)

Abstract

This paper proposes a new feature extraction method for speaker recognition, called Delta-Gated Spike Encoding (DGSE). The proposed approach combines log-Mel spectrograms, temporal delta, adaptive thresholding, an energy gate, and positive/negative spike encoding stages. The goal is to extract informative time-frequency changes from the acoustic signal and form a sparse representation suitable for subsequent spiking or energy-efficient models. Experiments were conducted on the VoxCeleb1 dataset using a three-stage parametric search procedure. The results of the initial coarse search, the subsequent extended search, and the final fine search were compared. The best result was obtained with the parameters alpha = 1.0, beta = 0.05, and energy_thr = -5.25. Under these settings, the method achieved total spike rate = 0.079585, sparsity = 0.920415, and gate open rate = 0.718533. The obtained results show that the DGSE method preserves informative parts of the signal while ensuring high sparsity. This makes it a promising solution for resource-constrained devices and CNN-SNN hybrid speaker recognition systems.