A Self-Supervised Representation Learning Method of Speech Recognition for Smart Grid

Recently, speech recognition has rapidly become a hot topic of machine learning and is widely applied in various domains. Despite impressive performance, speech recognition is still confronted with great challenges from different environmental factors in real world. To solve the challenges, a novel speech pattern recognition method for the pattern recognition task on speech data is proposed in the paper, which effectively combines machine learning techniques with acoustic model analysis. Specifically, the input speech data are first split into fragments, and then a clustering representation encoder network is utilized to extract the context representations of input speech data followed an aggregation module capturing clustering patterns. In the proposed method, the influence of environmental factors can be reduced as far as possible, and the robustness can be guaranteed effectively. Experimental results demonstrate the superior performance of the proposed method.