Suppression Algorithm Optimization and Real-Time Speech System Development based on Speech Recognition Technology
Fuxiang Du, Xianyan Yang · 2025
To address the problems of low recognition accuracy and poor real-time performance caused by noise in real-time speech recognition systems, this paper designs a noise suppression optimization algorithm based on a DNN-GRU combined neural network based on a deep neural network (DNN) and a gated recurrent unit (GRU), and applies the Dropout regularization method for optimization. Firstly, a one-layer gated recurrent unit (GRU) is combined with a three-layer deep neural network (DNN) to construct a DNN-GRU combined neural network model in order to fully utilize the temporal memory capability of GRU and the feature extraction advantages of DNN. Next, in the optimization of the noise suppression algorithm, the noise characteristics are learned by using the DNN-GRU model. The model obtains the amplitude characteristics of the pure audio in the training phase and is used to enhance the pure components of the input audio in the testing phase. Dropout regularization is introduced during the training process to randomly discard some neurons to suppress the risk of overfitting. Experimental results show that the recognition accuracy of the algorithm on the LibriSpeech and AISHELL-1 data sets reaches 96.8% and 95.7% respectively. In terms of anti-noise performance, the signal-to-noise ratio (SNR) is significantly better than traditional DNN and CNN methods. The optimized noise suppression algorithm provides new ideas and technical support for building a more robust and efficient real-time speech recognition system.