| With the development of audio event recognition, algorithm architecture becomesmore and more complex, the number of parameters increases. So the difficulty and timeconsuming of parameter optimization increase significantly. A method to optimizeparameter of model training level algorithmand a multi-level parameter optimizationmethod (Feature extraction level, Feature vector generation level, Model training level) areproposed to solve the problem of parameter optimization, decrease the optimization trainingtime, and improve the recognition accuracy.A method to optimize parameter of AdaBoost algorithm is proposed to solve theproblem of parameter optimization of model training level algorithm, and relieve theproblem of time consuming of grid optimization method. The proposed method analyzes theparameter of AdaBoost algorithm, confirms the dynamic pheromone update strategy basedon ant colony algorithm, then selects the proper ant number, and gets the optimal solutionfast. The experimental results show, the proposed method can improve the optimizationspeed by18times under the conditions of better recognition accuracy.To achieve parameter optimization of multi-level of audio event recognition, a methodis proposed to solve the problem of system overall parameter optimization. The methoddoes binary cascade coding to six parameters of feature extraction level, feature vectorgeneration level, and model training level, then selects the proper fitness function andTermination condition, and gets the optimal solution of the parameters. The experimentalresults show, the proposed method can improve the optimization speed by2×106times, F is89.29%, the accuracy rate based on evaluation method of segments is88.9%, the recall rateis93.15%.Designed and implemented a parameter optimization prototype system. Combinedwith audio event recognition algorithms, the system is modular in design, flexible ofparameters configuration according to the demand of experiment, easy to apply and has astrong scalability. |
