Research On Speech Enhancement Algorithm Based On Deep Neural Network

With the development of science and technology, speech has gradually become an important channel of human-computer interaction. However, in the reality environment, speech signals are inevitably affected by various kinds of noise, which generates the delay and misrecognition of human-computer interaction. Therefore, the emergence of speech enhancement technology is particularly important. This technology can eliminate the noise and improve the performance of the speech signal, which makes the speech clearer, more intelligible and more fluent.This paper mainly focuses on spectral subtraction and neural network speech enhancement algorithms. The former is a classical speech enhancement algorithm, which is easy to be implemented and understood. However, the enhanced speech has noise residue, namely, musical noise which is hard to eliminate. This thesis aims to use the latter way deep neural network algorithm to solve this problem.Deep neural speech enhancement algorithm is to enhance speech by training the network with the noisy speech, then the network has the ability to resist noise, thus to achieve the purpose of speech enhancement. In the training process, the BP algorithm is generally used. However, with the increase of the number of hidden layers of the network, the local optimization problem and the over fitting problem are easy to occur. Therefore, in this paper, layer greedy unsupervised training algorithm is implemented. The whole training process involves two parts---unsupervised training and supervised training. In the training process, an important parameter is the number of hidden layers of the network, which directly affects the result of network training. In this paper, the deviation method is proposed to set the number of hidden layers of the network. The appropriate number of hidden layers of the network is set according to a theoretical derivation, and it is tested in the experiments. However, the requirement of database is very high in the training process of deep neural network speech enhancement, and the time of training is very long. In order to solve these problems, a speech enhancement algorithm based on the combination of spectral subtraction and deep neural network is proposed in this paper.This combination algorithm solves the problem where the database requires many kinds of noise in the training process of deep neural network speech enhancement algorithm. Experiments prove that this combination algorithm is better than other combination algorithm of spectral subtraction.This paper proposes that deviation method is used to set the number of hidden layers of deep neural network, and at the same time, a speech enhancement algorithm based on spectral subtraction and deep neural network is proposed. Finally, the experiments on the above methods are verified, the result of experiments show that these methods are effective, and the deep neural network speech enhancement algorithm has a good effect of speech enhancement.