| With the development of the current level of science and technology,voice information interaction technology has been extended from human-human interaction to human-computer interaction,and more and more voice devices and applications are closely related to People’s Daily life.In real life,speech information will inevitably be polluted by a variety of complex environmental noise in the process of transmission,which seriously affects the transmission and use of speech signals,so it is indispensable to use speech enhancement technology to improve the quality of speech.Compared with the traditional signal processing methods,the speech enhancement method based on neural network has better enhancement performance in both stationary noise signal and non-stationary noise signal.The neural network method is relatively easy to integrate complex learning objectives,which greatly promotes the development of speech enhancement technology.However,most of the current neural network algorithms with excellent enhancement performance have problems such as large scale,high algorithm complexity and difficult training.In addition,most of the neural networks combined with traditional speech enhancement algorithms have limited enhancement performance of small models with low complexity.To solve these problems,this paper proposes a low complexity speech enhancement algorithm based on neural network.The main work is as follows:(1)A low complexity speech enhancement algorithm based on SAENN is proposed.In this paper,the relevant characteristics and extraction methods of the Meir spectral features of speech signals are analyzed and introduced in detail.Then,the extracted low dimension Meir subband features are taken as the input,and the subband amplitude estimation neural network SAENN is used to model and predict them to realize the enhancement of noisy speech signals.The selection of sub-band features effectively reduces the dimension of the network layer and ensures that the overall algorithm complexity can be kept at a low level.The Mayer spectrum features contain some characteristics that are in line with the human hearing perception mechanism,which effectively guarantees the SAENN speech enhancement algorithm’s good modeling and prediction ability in the model structure of low complexity,so that the algorithm can obtain good speech enhancement effect.(2)A low-complexity speech enhancement method based on TSRNN is proposed.Aiming at the fact that the proposed SAENN speech enhancement algorithm does not introduce phase information,which makes the algorithm limited in speech enhancement performance,a phase compensation algorithm based on PSONN and SAENN parallel cascade form a two-stage recurrent neural network enhancement system to further improve the performance of low-complexity speech enhancement algorithm.PSONN is mainly used to estimate the noise amplitude required by the phase compensation algorithm,with the phase of the enhanced speech after PSC,SAENN is used to predict the amplitude of the enhanced voice,and the two are processed in parallel to achieve speech enhancement,and the enhanced voice of the algorithm system also obtains a better objective evaluation score. |
