Research On Language Recognition For Russian Military Speech

With the continuous advancement of various means of information acquisition,the information obtained becomes more,easier and more redundant,as well as voice information.With the emergence of more and more multilingual speech environments,the elimination of all redundant information in non-target languages in speech information becomes more and more critical,and the need for speech recognition for speech is also growing.In order to fill the gap in the current domestic language recognition research for Russian,this paper will combine the deep learning method to study the language recognition of Russian military speech.In this paper,by studying the characteristics of Russian pronunciation,it is found that the speech spectrum diagram contains language discrimination features,and through comparative experiments,it has been proved that the spectrum diagram feature has a good recognition effect on Russian speech in the military field.Besides,the data set is divided by cross-validation method,and two comparison experiments are carried out.Firstly,the recognition performance of different language recognition features and classification models are explored through comparative experiments.In this process,three feature extraction methods,such as phoneme sequence feature extraction,spectral feature extraction and Gaussian modeling feature extraction,are studied.The full-space modeling identification vector ivector based on factor analysis obtained by Gaussian modeling feature extraction method is analyzed.Characteristics;Convolutional neural networks capable of capturing image features in deep learning methods are also studied.The performances of traditional language recognition methods and CNN-based deep learning language recognition methods in Chinese,Vietnamese,Russian,Spanish and Japanese languages are compared.difference.Secondly,by synthesizing the Russian military speech corpus,the language recognition model for Russian military speech is trained to compare ivector features and spectrogram features.In this process,the characteristics of Russian military speech corpus are analyzed,and the influence of ivector dimension on recognition performance is studied.Through experiments,it is found that the recognition accuracy of the CNN-based language recognition method is significantly improved compared with the traditional language recognition method on the five-language recognition task;Based on the spectrogram feature-CNN,the language recognition method performs best on the recognition task of Russian military speech corpus,and obtains 100% recall rate and 99.2% accuracy rate for Russian military speech corpus.On this basis,this paper constructs a language recognition prototype system that can accurately and quickly extract Russian military speech from various voice information in a specific environment.The prototype system is capable of recognizing Russian military speech with a recognition accuracy of 99.8%.The experimental results show that the language recognition based on deep learning method is obviously superior to the traditional language recognition method.And the ivector feature based on Russian pronunciation features does perform better on Russian recognition tasks.Besides,using deep learning based on Russian military speech corpus as a front-end part of speech data processing,the language recognition prototype system can improve the extraction efficiency of Russian military speech.The effect of the deep learning method is closely related to the quality of the data set.The prototype system is not good for other environmental speech recognition,and the speech of the specific environment is used for training again to enable the neural network to better recognize the speech in such environment.