A Study Of Sound Recognition Algorithms Based On Neural Network

With the advent of the era of big data, in both fields of industrial production and normal life are filled with a lot of multimedia data. Sound, as an important part of the multimedia data contains a lot of information. Sound signal processing and analysis can extract a lot of useful information from the data, so for sound processing and analysis technology has always been a study hot-pot for researchers all around world. In recent years, sound recognition technology has also obtained a lot of attention and applications. Sound recognition which compares the sound features to be identified and the sound samples, so to reach the judgment of consistency of them. Sound recognition can be applied in many fields and occasions, such as abnormal background sound detection, audio information retrieval, and audio media copyrights detection.Prior to the recognition stage, the sound needs to be pre-processed. Sound recognition pre-processing includes pre-emphasis, framing and windowing, and endpoint detection. On the basis of the pre-treatment of the sound, the technology of feature extraction are introduced to obtain the feature vectors of the sound. Then the pattern matching stage provides the final result of the sound recognition.Based on the basic principles of the neural network, this paper mainly studies the methods of applying neural networks to solve the pattern match problems in multi-class sound recognition. The main work in this paper contains:Firstly, based on the basis of neural network, the specific parameters of two-class neural networks are discussed are obtained, such as the transfer function, the number of neurons and nerve layers are determined. Secondly, the sound recognition schemes are demonstrated and compared. We provide three recognition schemas, they are the linear identification schema, the parallel ranking schema, and the two-class promotion schema. After compared, the two-class promotion schema is determined to be the basic method of multi-class sound recognition program.Thirdly, the methods of using two-class neural networks to finish the multi-class sound recognition are elaborated. The multiple-set-compete method and the algorithm of result reliability are fully demonstrated. Multiple-set-compete method can significantly improves the recognition rate of two-class sound recognition. Specific experimental results verifies the utility of the multiple-set-compete method in multi-class recognition. The algorithm of the result reliability applied to the recognition program allows users to grasp the initiative in the recognition procedure, to get the satisfactory recognition result balanced between the recognition time and the recognition rate.Finally, aiming at the multi-class of any number sound recognition, the group match competition method is introduced. A number of specific examples are discussed, and the multi-class sound recognition within10classes are given recommended grouping models. Problems with a greater number can be solved by firstly grouped into groups of ten or less.