| Aviation noise not only seriously affects the physical and mental health of civil aviation personnel,but also seriously pollutes the speech communication between them,leading to blurred or even erroneous communication between both parties.In some cases,the time consumed and incorrect speech information required to repeatedly confirm due to blurred communication may lead to very serious consequences,even endangering flight safety.Currently,the speech enhancement technology used in the field of civil aviation is mainly based on traditional unsupervised algorithms.Due to the limitations of these algorithms,the effect of speech enhancement is not ideal when dealing with aviation noise with high sound level and non-stationary characteristics.In order to carry out speech enhancement processing for noisy speech in aviation noise,a speech enhancement method based on improved Conv-Tas Net neural network is proposed in this thesis.Firstly,in view of the relatively complex composition of aviation noise,this thesis collects the aviation noise generated by the most common aircraft types in the domestic classic airports,such as the A320,Boeing 737,and ARJ21,under four flight postures of takeoff,landing,circling,and taxiing,through independent recording and online collection;For pure speech samples,this thesis selects TIMIT and AISHELL-1 public data sets;The samples are firstly filtered,cropped,expanded,and processed at equal sampling rate and bit depth.Five different signal-to-noise ratios of 10 d B,5d B,0d B,-5d B,and-10 d B are synthesized to obtain noisy speech samples by additive superposition,and the final sample set is established.Secondly,for the characteristics of aviation noise with high sound level and non-stationary characteristics,it leads to the problem that the baseline Conv-Tas Net network model has serious loss of features of speech signals with low signal-to-noise ratio and inaccurate prediction of non-stationary noise when performing speech enhancement with noise in aviation noise.In this thesis,we incorporate the Multi-head Attention mechanism and the improved ECA-Net mechanism into the Conv-Tas Net network.The ablation experiment results show that,under five signal-to-noise ratios,the network model incorporating the two mechanisms improves speech enhancement performance.Compared with baseline Conv-Tas Net,the Conv-Tas Net,which integrated both the Multi-head Attention mechanism and the improved ECA-Net has an average increase of 0.131 in MOS score,1.8128 d B in seg SNR,0.0771 in PESQ and 0.0413 d B in STOI.Then,in order to verify the real-time performance,the proposed algorithm and the DNN neural network model based on time-frequency analysis are analyzed for model parameters and processing time.The experimental results show that,the proposed method has prominent advantages over DNN neural networks in real-time performance.Finally,in order to verify the generalization ability,the proposed algorithm in this thesis is compared with three traditional unsupervised speech enhancement algorithms,log MMSE,Spectral Subtraction,and Wiener Filtering,to perform experiments on real air-to-ground conversations.By observing the Mel spectrogram,it can be clearly seen that the performance advantages of methods based on neural network for speech enhancement processing in aviation noise are better than traditional unsupervised speech enhancement algorithms.The algorithm proposed in this thesis has a good generalization ability. |
