Research On The Evaluation Method Of Speech Communication Interference Effect

The evaluation of speech interference effect refers to the technology of analyzing the interfered speech signal received by the communication system to determine speech interference effect,namely the impairment level.In the environment of electronic countermeasures,countermeasures will interfere with communication equipment to reduce our communication effectiveness.At the same time,with the popularization of 5G communication standards and the advent of the Internet of Things,the number of communication devices has increased rapidly,and unintentional interference has also increased,which has an unpredictable impact on the performance of communication systems.Accurately evaluating the interference effect is of great significance to the development of communication countermeasure equipment,the assessment of the situation of electronic countermeasures and understanding of communication quality.This paper conducts research on the evaluation of communication interference effect in the presence of strong interference.The specific research content is as follows:Firstly,this paper models the problem of evaluating the speech interference effect,and studies the reference evaluation method based on traditional machine learning.Based on three features proposed based on prior knowledge,a dynamic time wrapping algorithm is used to calculate the speech interference measure.Then it studies the evaluation method based on a single measure and the evaluation method based on the fusion of multiple measures.Based on the measured data,the two methods are compared and analyzed.Experiments show that the multi-metric fusion method has the best performance.The correlation coefficient can reach 0.97 on the indoor dataset,and the RMSE can reach 0.36.Secondly,this paper studies the evaluation method of communication interference effect based on deep learning.First,the disturbed speech is converted into the image representation of the log-mel spectrogram.After the dataset of the spectrogram is obtained,a convolutional neural network model is constructed.And features are automatically extracted from the log-mel spectrogram,and classified,and no reference evaluation is realized.The effectiveness of the method is verified by the measured data and compared with the multi-measure fusion method which has the best performance among the evaluation methods based on traditional machine learning.Experimental results show that this method can accurately determine the impairment level.On the indoor dataset,the classification accuracy can reach 87%,which is better than the multi-measure fusion evaluation method.Thirdly,aiming at the problem of insufficient training data for disturbed speech in a specific scenario,this paper studies the interference effect evaluation method combined with transfer learning.Based on the convolutional neural network model that evaluates the spectrogram,first,this article uses two strategies to fine-tune the pre-trained model to improve the performance of the target task.Then use the method of maximizing domain confusion to train a domain adaptation model that can simultaneously learn domain invariant features from the source domain and target domain.The experimental results show that both methods can achieve improvement on the target task,and the domain adaptation model has a significant improvement in accuracy of 10% compared with the best baseline model.Finally,this paper investigates the real-time interference effect evaluation method for the problem of estimating the interference effect simultaneously during the communication process and to continuously correct the estimation with the accumulation of data.First,an incremental evaluation network capable of multiple estimations and gradually adjusting the estimation results is designed.Then two different objective functions are introduced,and the algorithm is verified on indoor dataset,and the model performance is compared for different objective functions,and for different number of sub-network.The results show that the estimated accuracy rate can reach 74% at 1/3 time of communication,and with the accumulation of data,at 2/3 time,the accuracy can reach 85%,and at the end of communication,it can reach 91%Accuracy.