New Audio Event Detection Based On Atomic Model

With the rapid development of Internet technology and the popularity of handheld recording equipment,there is an explosive growth of audio data easily accessible by people.In order to manage and make use of the massive audio data effectively,the research of audio event detection technology has become a hot spot.The current audio detection algorithm is mainly for pre-defined audio types,while the new audio event discovery is relatively lacking.The new audio event defined in this article refers to a class of audio events that may be of interest to an unknown audio stream,the number of which is greater than other types of audio events over a period of time.Since the type of the new audio event is known only after it is detected,the new audio event is found to be unsupervised.In this paper,we mainly improve the traditional vector quantization method,propose a new audio event detection method based on atomic mode,and verify its performance through experiments:1,In the codebook training stage,we use a normal distribution to represent a codeword,combined with the idea of fuzzy clustering.So when in the stage of codebook training division,a sample can be subordinate to more than one codeword,to achieve a "soft classification" Quantization error.2,In the codeword searching stage,a coding method based on dynamic programming is proposed,considering the quasi-stationary characteristic of the target audio event.This method smoothes the encoding of each segment of audio,reduces the influence of background noise such as impact noise,and improves the stability of the system in complex audio environment.3,In the codeword index stage,the codewords are called as "high frequency codeword",which containing most of the samples.We count the number of frames in each segment of the audio belonging to the high frequency codeword,and use this as a basis for the screening of new audio events.4,Design and implemented a new audio event detection system.Experiments show that when the target audio type accounted for more than 20%,or the target audio duration is more than three times the size of the 2ndone,the system can generally achieve the goal ofnew audio discovery.The extraction of the target audio samples also achieved good results.In summary,this method provides a solution to discover new audio events in the absence of a priori knowledge of the audio set,which makes up for the lack of research work in this scenario.