Research On Abnormal Audio Recognition Algorithm Based On MFCC And GMM

As one of the audio surveillance system, abnormal audio recognition is the process of automatically recognizing which is based on the features included in abnormal sound waveform. Because of its particularly advantage on high efficiency, economy, small complexity and protection of people's privacy, this technique can be combined with video surveillance system.Therefore, the audio surveillance technology is very promising, and it is worth a lot of scientific researchers to engage in their studies.To overcome the problem of low rate rate and high complexity in abnormal audio recognition, the abnormal audio recognition system based on the Mel frequency cepstrum coefficients and short-term energy is proposed. This feature vector makes the Gaussian mixture model classifier available than the use of MFCC and Differential MFCC features for better classification performance.The classifier can achieve an average recognition rate of more than 90%, and the small computational complexity.It is showed the elaborate steps of system implemention, and proved the effectiveness of the algorithm with results from simulation evironment.In the aspect of performance research, The author analyzes the recognition rate of different features, in which the recognition rate related to the features performance. The author also tests the performance of different numbers of Gaussian mixtures, in which the choice of mixture numbers related to the training data are concluded.The author analyzes the EM and MDL algorithm in which the MDL algorithm has a better use of space.In aspect of system construction, this author describes the implementation of full abnormal audio recognition system by MATLAB, including audio signal preprocessing, feature extracting and classifier training and recognition. In the preprocessing module, the original audio is normalized, pre-emphasis, overlapping divide; In feature extraction module, It uses MFCC, ZCR, Energy, LFCC as feature parameters. In the model training and recognition module, author use the Gaussian mixture model as a classifier, which in comparison with other classifiers, can better identify abnormal audio difference. The audio abnormal audio recognition system classify and discriminate eight abnormal audio. The system is one of the best system because of its high recognition rate and good performance, simpleness and high efficiency, privacy.