Research On Embedded Robust Speech Recognition Based On Support Vector Machine

The robustness issue of the embedded speech recognition in the noiseenvironments is the key point to achieve the speech recognition technology beingapplied to the real-world applications from the laboratory research theory. However,without considering the application object’s processing capacity, most of Noise-robustresearches, with complex computing algorithms, are difficult to be applied to theembedded speech recognition system now. To solve this problem, this paperresearches on the speech enhancement theory, support vector machine (SVM) theoryand embedded speech recognition theory, forms a methodology of robust speechrecognition which is suitable for the embedded system, speeds up the pace of thespeech recognition technology applied to the real-world applications from laboratoryresearch theory.Firstly, this paper sets up the Mathematical model for speech signal by the speechsignal digital processing theory, and then researches on the speech enhancementmethods based on the above model, especially the Kalman filtering algorithm. Thetraditional Kalman filtering algorithm for speech enhancement needs to calculate theparameters of AR (auto-regressive) model, and to perform a lot of matrix operations,which usually is non-adaptive. The speech enhancement algorithm is proposed in thispaper eliminates the matrix operations and reduces the calculating time by onlyconstantly updating the first value of state vector. Design a forgetting factor foradaptive filtering, to automatically amend the estimation of environmental noise bythe observation data, so that the algorithm has stronger robustness.Secondly, the statistical theory and the SVM theory are established in detail.Based on both of these and the characteristics of the embedded speech recognition,this paper researches on the improvement of the key technologies which SVM isapplied in the embedded speech recognition: solves the problem caused by thevariability of duration of speech utterances which the traditional SVM can’t handle,by embedding DTW which has good time dynamic programming ability into theGauss kernel function; Analysis to Gaussian kernel function, we design a simple andeffective method for the selection of SVM’s parameter by the characteristics of theGaussian kernel function; According to the quadratic programming theory, we focuson the SMO algorithm which is used as the training algorithm for SVM, use theKeerthi’s working set selection method to shorten the system’s training time, and usen linear arrays instead of n nkernel matrix as the cache mechanism to reduce thememory; Review on the multiclass classification SVM in speech recognition,analyze the multiclass classification algorithm by the graph theory, we use the Directed Acyclic Graph (DAG) which has more stronger generalization ability andjust needs to do the decision function k-1times for classification in the system’stesting stage. Compare to the "one to one" method, DAG has higher recognition rateand less testing time, and there will not be misclassification and unclassifiable region.Finally, based on the above researches, this paper presents a speech recognitionsystem on the Matlab and shows the detailed numerical results.