DNN Regression Model and Microphone Array for Noise Reduction System on Real Field

Abstract

In this study, we proposed a regression model based on the deep neural network (DNN) and a microphone array for noise reduction system on a supermarket robot. The proposed speech enhancement model can be separated into two parts: the training phase and the enhancement phase. A large amount of data is used for DNN model to train the mapping function, and to separate the estimated features of clean speech signal according to the features of noisy speech signal. In addition, convolutional recurrent neural networks are added to enhance the output performance of the original DNN model. The main objective is to take advantage of convolutional networks in feature extraction and to use recurrent networks to have the ability to process temporal models. This proposed system is used to overcome the noise interference in real life. The system can achieve the better speech recognition rate and provide correct sentences for the backend dialogue system. The microphone array is used for collecting audio, which not only to reduce the noise at the front end, but also to enhance the collected audio distance of the system.