A Hybrid Approch for Noise Suppression Based on Recurrent Neural Network

A Hybrid Approch for Noise Suppression Based on Recurrent Neural Network

연구 분야: Artificial Intelligence

학회: 2023 8th International Conference on Intelligent Computing and Signal Processing (ICSP)

Neural network is widely used in machine learning, aiming at the problem that the traditional noise suppression methods are not accurate enough in noise recognition. This article introduces a speech enhancement method based on the principle of spectral subtraction using the recurrent neural network (RNN) model. In the feature extraction process, Bark spectral coefficients are used so that reducing numbers of the features of the neural network model to reduce the computational complexity. The three steps in spectral subtraction are replaced with RNN model to improve recognition accuracy, constructing the gated recurrent unit (GRU) model by introducing update and reset gates into the RNN. Synthetic speech data and collected speech data are used for training to determine the parameters of each layer of the network and avoid problems such as vanishing gradients. The model's effectiveness is analyzed in test scenarios of canteens and convenience store with background noise on campus, and the RNN model shows more significant denoising effects compared to traditional spectral subtraction methods.