The Algorithm Study For Continuous Speech Recognition Based On Convolutional Neural Network

At present,speech recognition is increasingly becoming an essential function of the products in the high-quality service industry.Therefore,the accuracy and efficiency of recognition have become the key to application.Research in industry shows that there is a direct relationship between the efficiency of recognition and the level about efficiency of training.The main reasons for the efficiency of training lie in the redundancy calculation and low degree of fitting caused by whether the adaptive adjustment of the weights in the acoustic model can fully meet the change of error.To improve the accuracy of speech recognition,it is the key to increase the noise for data set and improve method of endpoint detection.Combining research achievements at home and abroad,analyze and study the difference of characteristics between speech and noise to enhance short-term energy and improve the sensitivity of threshold decision;increase noise on differential data sets to enhance the robustness of recognition.By improving the back propagation algorithm to constrain the range of weight,avoid oscillations and shorten the time of training.Finally,build a prototype system for speech recognition to verify the effectiveness of the algorithm.The main work of this article is as follows:(1)A double-threshold method of endpoint detection to enhance short-term energy and a method of adding noise in differential data set are proposed.Because the randomness of the background noise cause the low accuracy of the endpoint detection and the accuracy for speech recognition is low in a certain environment,this paper analyzes the similarities and differences between the short-term energy and the cosine of the autocorrelation function.The short-term energy of the speech segment obtained by the cosine of the autocorrelation function.Compare the short-term energy of the effective speech with the cosine of the autocorrelation function to achieve the purpose of enhancing the short-term energy and enhance the ability that threshold can determine the position of the endpoint.From the reverse of subtraction about spectrum,the background noise with environmental specificity is added to the speech of the training set after the endpoint detection.Compensate the training set by the way of spectrum domain can reduce the difference between the training set and the application environment and improve the robustness of recognizing the noisy speech.(2)Propose a back-propagation algorithm of narrowing weight range(NWBP).In the system of real-speech recognition,there is problem such as low efficiency of training causedby massive training data and hyper-scale parameters of convolutional neural network.For these problems,the NWBP algorithm uses the K-MEANS algorithm to obtain the seed node that approaches the minimum value of error when searching for the minimum error around the end of training parameter and causing the phenomenon of easy-to-occur oscillation.It avoids oscillations by reducing the range of weights with the principle of the boundary to make the error of network converge as quickly as possible and improve the efficiency of training.Through simulation experiments,the convergence of NWBP algorithm compared to the other algorithm is improved in the process of training weight for complex convolutional neural networks.To some extent,the algorithm reduces the redundant calculation,shortens the time of training,and can better reflect the advantages of accelerating convergence.(3)Build a prototype system of speech recognition.Based on the SRILM tool for training the model of language and the PocketSphinx tool for decoder,the module functions of the prototype system are designed and implemented,and the effectiveness of the proposed algorithm is verified by using corpus in different environments.