With the continuous development of science and technology,People's needs are also getting higher and higher,We urgently hope that the communication between people and computers will be intelligent,that is,the machines understand human language,understand the orders of human beings,to achieve human-machine interaction,let the machine become more intelligent and more human,which prompted researchers to strengthen the understanding and recognition of speech recognition technology,thus promoting the rapid development of speech recognition technology.After more than 60 years' efforts,speech recognition technology has made some achievements,but there are still some technical problems in the humanization of communication between humans and computers.Researchers need more in-depth research.In quiet environment,the recognition rate of speech recognition system is high and robust,but in noisy environment,the recognition rate and robustness of speech recognition system will be seriously affected.Therefore,the research of speech recognition system's high recognition rate and robustness in noisy environment is of great significance.The basic principle of speech recognition technology,the key steps of speech recognition technology and the basic frame of speech recognition system are introduced in this paper.First of all,the whole speech recognition system is analyzed,includes speech information preprocessing,speech information feature extraction,speech information model training,speech model parameter library and pattern matching.Then,we focus on the feature extraction part of speech information,the process and implementation steps of LPCC feature parameter extraction algorithm,MFCC characteristic parameter extraction algorithm and PNCC feature parameter extraction algorithm are detailed deduced and analyzed.The PNCC feature extraction algorithm is inspired by the heuristic algorithm of MFCC feature extraction,it can be seen as an improved MFCC feature extraction algorithm.Compared with MFCC,the PNCC parameter uses a Gammatone filter group to replace the trigonometric filter group in the process of MFCC to simulate the characteristics of the human ear basement membrane;in addition,PNCC uses the nonlinear power function and the continuous power deviation instead of the logarithmic operation in the process of MFCC parameter extraction and calculation to reduce the impact of background noise,it is more consistent with the compression perception of the auditory nerve of the human ear.But,MFCC and PNCC all need to be Fourier transform to get the speech signal spectrum information,the Fourier transform is due to the simpleness of the resolution change,there is a natural defect in the processing of nonstationary signals for speech,it can only obtain a general signal which contains frequency components,There is no knowledge of the moment of the composition,as an effective method for processing nonstationary time series signals,wavelet transform has the characteristics of time-frequency two domain multi-resolution analysis.Therefore,this paper combines wavelet transform and PNCC speech features to propose two new characteristic parameters W-PNCC and WP-PNCC,the extraction process is the introduction of wavelet transform and wavelet packet transform in the front processing stage of PNCC speech parameter extraction.The speech signal is refined by scaling and translation,and the speech signal features are detected and analyzed in different scales.On the basis of wavelet transform and PNCC speech feature extraction,a speech recognition algorithm based on BP neural network is proposed in this paper,and the basic process of the algorithm is described.The standard BP neural network model is introduced,and its advantages are analyzed.The extracted speech feature parameters are input into the neural network model,Through continuous learning and training,the recognition model of different model parameters is constructed.By comparing the recognition results,the network model with the best recognition effect is selected,and the characteristic analysis and contrast experiment is completed by using the network model.Finally,a lot of comparative experiments and performance analysis of several speech feature extraction algorithms and their recognition performance are carried out through MATLAB simulation software.According to the influence of the noise and recognition rate of the speech recognition system,the feature extraction algorithm is compared and analyzed.Experimental results show that the improved feature parameters proposed in this paper not only inherit the robustness of the original PNCC feature parameters,but also improve the recognition rate of the system,which is better than the original PNCC feature parameters,and is more robust against background noise.Therefore,the speech feature extraction algorithm proposed in this paper is more suitable for the speech recognition system,and it can serve the practical application better. |