Research On Speech Recognition Technology Based On BP Neural Network

Since the speech recognition technology has broad application prospects, it has received high attention from people. The dynamic time warping (DTW) and Hidden Markov Model (HMM) technologies are widely used in the current research of speech recognition, and the relevant research has achieved a mature theoretical basis. The voice is not a simple linear process but a complex nonlinear process, so the artificial neural network with virtues of non-linear, adaptive and learning has been applied to the speech recognition system in recent years and good results are obtained. The standard three-layers BP neural network is taken as a recognition algorithm to explore its application in the speech recognition technology in this paper.System research on the isolated numeral recognition using BP neural network algorithm and experimental study on the neural network structure and model parameters is conducted in the paper. The recognition rate of speech recognition system is improved, and a specific person isolated word speech recognition system is ultimately realized. This paves the way for the subsequent non-specific speech research.The basic theory of speech recognition is introduced in detail. The speech signal preprocessing is analyzed, including voice signal acquisition, filtering, windowing, framing, endpoint detection. The important research on the feature parameter extraction methods of LPCC and MFCC are made, and the later method is improved, and hybrid MFCC parameter extraction method is proposed. Experimental analysis on the LPCC and MFCC hybrid parameters and LPCC and improved MFCC hybrid parameters are carried out. The recognition performance of different parameters are compared with each other and results show that improved MFCC parameters and hybrid feature parameters have better recognition performance. Moreover, the basic principle of BP neural network, learning rule and algorithmic process are introduced and analyzed, and the combination of increased momentum factor method with adaptive rate method is presented. This algorithm effectively solves the problems of the neural network which is easy to fall into local minimum point and has a slow convergence rate, and it greatly optimize the system performance. Finally, the effect of different numbers of nerve cells and training samples on the recognition accuracy is investigated. Results reveal that the numbers of nerve cells and training samples have great influence on the recognition accuracy. Therefore, the numbers should be chosen appropriately in the experiment.