The Research Of Feature Extraction Algorithm On The Speaker-Independent Speech Recognition

With the rapid development of artificial intelligence,speech recognition has become the key technology for human-machine dialogue.The technology of non-specific people's speech recognition has always been one of the focuses of people's research.At present,there are still many difficulties,such as inaccurate extraction of speech feature and the low accuracy of the detection under low signal and noise conditions.Aiming at the extraction of eigenvalues of non-specific human speech recognition,an improved algorithm for extracting two eigenvalues is proposed.The basic process of speech recognition can be divided into three steps including speech signal preprocessing,feature parameter extraction,and the model of speech training and recognizing.Firstly,the development history and key technologies of speech recognition in recent years are elaborated and analyzed.Secondly,a pitch detection algorithm based on lifting-wavelet-weighted linear prediction autocorrelation function is proposed to solve the problem that the pitch extraction algorithm is not effective under low SNR conditions.Experimental results show that the detection accuracy and robustness of the proposed algorithm are superior to other algorithms under different SNR conditions.Finally,aiming at the defects of Mel frequency cepstrum coefficient and linear prediction cepstrum coefficient,a feature extraction algorithm of MFCC fusion pitch period based on empirical mode decomposition is proposed.The experimental results show that the corresponding improved algorithms compared with the traditional cepstrum feature parameters can contain more semantic information,which makes the characteristic parameters of speech word higher.In the thesis,a non-specific human speech recognition system based on hidden Markov identification model is built under the MATLAB platform.The recognition results of the improved algorithm in the thesis are compared with the recognition results of other feature parameter extraction algorithms,highlighting the recognition performance.Finally,according to the application requirements,an interface with interaction of human and machine is designed to improve the system performance.