Feature extraction of speech signals of exoskeleton devices in noise environments

In actual working environments,exoskeleton devices for speech systems have poor voice command recognition performance due to the influence of environmental noise. This paper presents speech characteristics based on the Gammatone Frequency Cepstrum Coefficient using discrete orthogonal Stockwell transform. Time domain information of speech signal energy and zero crossing rate is characterized by discrete path transformation and composed into hybrid features. Redundancy,irrelevance, and information complementarity between the features are considered under low signal-to-noise ratios. The improved correlation fast filtering feature selection algorithm is used to obtain the optimal feature subset for the voice system of exoskeleton device control commands. Experimental results show that the optimized hybrid features are more robust under low signal-to-noise ratios,and the recognition rate of traditional Mel cepstral coefficients improves by about 20% under pink noise with zero signal-to-noise ratios. © 2024 Editorial Board of Jilin University. All rights reserved.