The Realization Of The Multi_type Audio Events Detection

Nowadays, along with the social development, human has begun to enter the information age. Audio is an important means of passing information, Audio data volume continues to grow.It Put forward urgent requirements for identifying, managing, using and sharing of audio data. And this depends largely on the detection and classification of audio events.Audio event detection is mainly divided into: statistical characteristic, characteristic model, template matching and deep learning approach. This paper puts forward a series of improved algorithm based on gaussian mixture model, and deeply discusses the performance of algorithm through experimental data.1, Improve the K Means clustering process based on the traditional GMM training method. Introducing the fuzzy function, this paper puts forward fuzzy clustering method,and make reasonable mathematical limit during sample separating at the same time.This makes a better model generation process. Experiments prove that the fuzzy clustering method is more reasonable compared with the traditional GMM model.The identification accuracy has certain ascend.2, Get the fuzzy gaussian code which used to quantify the audio data based on the idea of fuzzy clustering.The stirng compression algorithm in text processing is used to extract representative symbol string from audio quantitative string.We call these symbol strings as atomic patterns which can be a good characterization of typical characteristics of the audio type.3, Build joint GMM model using the atomic patterns of audio type.We treat a single atomic pattern as a local representative character of audio type and build a local GMM model to represent it.Multiple atomic patterns of the same audio type can build multiple local GMM models,we can represent the whole characteristics space through combining these multiple local GMM models.4, In this paper,we explore the performance of joint GMM model from two aspects that are the atomic structure and the atomic timing sequences. At last,we summarize that using the atomic model has good effect for accelerate large-scale model operation through experiment,it can greatly reduces the operation time.5, Based on the above algorithms,we build the audio recognition system whose architecture is based on MFC.Due to the audio volume is large while handling multi-type audio,some of the module uses multithreading..In the end,we make a summary about all the multi_type audio events detection algorithms based on GMM model,and put forward my own suggestions about next works or problems.