Research On Speech Enhancement And Robust Feature Extraction

With the development of global economic integration and computer technology,language communication from simple people and people has turned into an exchange between advanced people and computers.Human-computer interaction allows computer systems to recognize human speech and respond accordingly to human speech.The meaning of the voice responds accordingly.At the same time,in order to cope with the convenience and rapidity of modern human communication,the study of smartphone speech recognition systems that convert speech into text has become the focus of research in the field of speech recognition.The speech enhancement algorithm and speech feature extraction largely determine the "good" and "bad" of the speech recognition system.The main work of this paper is as follows.The speech enhancement algorithm in speech recognition front-end processing,for the traditional spectral subtraction speech enhancement algorithm can not follow the changes in the noise with the shortcomings of change,proposed improved LMS adaptive and self-encoding neural network speech enhancement algorithm.The improved LMS adaptive speech enhancement algorithm greatly improves the noise followability.For the insufficiency of the intractable noise reduction effect,a more prominent self-encoded neural network speech enhancement algorithm with universal applicability is proposed.Both of these algorithms achieve more effective noise removal.Disadvantages of the poor robustness of the feature parameters of the speech MFCC are Delta and Delta-Delta processing,and all the new feature vectors are synthesized.Then further linear discriminant analysis and maximum likelihood linear transformation(LDA+MLLT)and speaker are performed.Adaptive transform(abbreviated as SAT)obtains strong anti-interference speech features.The features are put into a speech recognition model for training and simulation respectively.Compared with the original feature,the speech recognition effect obtained after transformation is better,and realized the extraction of speech feature parameters.For the inability of wavelet transform to perform multi-layer decomposition of high frequencies,a voiceprint feature extraction algorithm based on wavelet packet transform that can perform multi-layer decomposition of low frequency and high frequency is proposed.The experimental comparison shows that the algorithm has a good recognition effect through experiments.