Environmental sound classification based on adding noise

As ESC is widely used in daily life, it has developed rapidly in recent years. Environment sound classification is a type of sound event recognition (SER). Because of the different position between the sound source and the physical medium of collecting information and the interference of many other sound sources in the process of receiving sound, it leads to the confusion and overlapping of sound events and other complex environmental sound. Combined with the characteristics of environmental sound mentioned above, neural network training with typical experience risk minimization is prone to memory of specific individual voice in the training stage, which will lead to unsatisfactory prediction results when predicting data outside the training distribution, that is, the occurrence of over-fitting. In order to solve the problem of low generalization of neural network, this paper started from the data source to explore the effect of adding gaussian white noise and SNR noise into environmental sound, which are two kinds of audio enhancement algorithms, and organized a series of experiments on Urbansoud8K public environmental sound dataset to verify.