Research On Intelligent Enhancement Technology Of High Frequency Speech Signal

Shortwave communication is widely used in military communications and emergency communications due to its outstanding advantages such as strong invulnerability,high mobility,long transmission distance and low cost.However,because shortwave communication mainly uses sky wave propagation,that is,the reflection of the ionosphere is used for transmission.Therefore,shortwave communication is susceptible to the influence of force majeure such as seasons,weather,and geographic location,resulting in obvious fading.At the same time,due to long-distance transmission loss,the quality of shortwave received voice signals is greatly reduced,and voice signals are an important form of shortwave communication.First,the low quality of shortwave speech signals has always been an urgent problem in practical applications.Speech enhancement is a technique used to suppress the background noise of speech signals.It is often used to improve the quality of speech signals.Traditional shortwave speech enhancement,such as spectral subtraction based on IMCRA(Improved Minima Controlled Recursive Averaging),uses the noise estimation results of the IMCRA algorithm to perform spectrum Reduction can achieve the suppression of background noise,but under low signal-to-noise ratio,because it is difficult to accurately estimate the noise,there are many residual background noises and the speech signal loss is serious.Deep learning has been widely used in the field of speech enhancement due to its powerful mapping capabilities,so this paper proposes to adopt a method based on deep learning to realize the intelligent enhancement technology of shortwave speech signals.Different from the problems faced by existing deep learning in the field of speech enhancement,shortwave speech signals are mainly in urgent need of solving the problems of low signalto-noise ratio and serious fading.Therefore,the intelligent enhancement technology of shortwave speech signals based on deep learning proposed in this paper is mainly Research around these two issues.Aiming at the background noise suppression of shortwave speech signals,this paper first uses Convolutional Neural Network(CNN)to implement a shortwave speech noise reduction technology based on deep learning.In order to further improve the noise reduction performance,this paper proposes the IMCRA-SS-IRM-CNN(CNN based on Ideal Ratio Mask and IMCRA based Spectral Subtraction,)method,which mainly uses the good noise reduction performance of IMCRA-SS and the powerful mapping of CNN Combining capabilities to enhance short-wave speech signals.Simulation experiments prove that this method improves the noise reduction performance and generalization performance of the convolutional neural network.Compared with the previous improvement,the average P of each improved method under each signal-to-noise ratio The ESQ score can be increased by about 0.1 points.Aiming at the fading problem of shortwave speech,this paper proposes two shortwave speech anti-fading technologies from the perspective of channel equalization and diversity reception.First,from the perspective of channel equalization,it is proposed to use E-CNN(Equalization CNN)and IMCRA-SS-IRM-CNN to perform blind equalization and noise suppression on shortwave speech,and compare the equalization neural network and noise reduction neural network through simulation experiments.The impact of different joint methods on the enhancement performance of shortwave speech signals;second,from the perspective of diversity reception,a method of using neural networks to first perform branch noise reduction,and then multi-channel diversity combination to achieve shortwave speech signal anti-fading.