The Research On Channel Adaptive Method Of Audio Event Recognition

For audio event recognition system which is in internet transmission environment,the difference of acquisition device, recording environment and encoding way all canproduce channel mismatch. The research of this paper mainly focus on how to resolvechannel mismatch problem introduced by coding difference and put forward the featuredomain and model domain channel adaptive method of audio event recognition.For the nonlinear mismatch problem introduced by audio coding difference, putforward a channel adaptive method of audio event recognition by feature mapping withadaptive matched Top-N weighted Gaussian components. This method is applied after thefeature extraction part and before the model training part. For each extracted feature data,use the Top-N scored Gaussian components of the corresponding channel model tocomplete linear weighting and feature mapping. The value of Top-N is adaptive-matchedobtained by score threshold, which can not only cover even channel information, but alsocan avoid the defects of poor generalization of fixed Top-N method for the channelmodels which have the different Gaussian components. Experimental results show thatthe proposed method has a better channel adaptive ability compared with the Top-1Gaussian component feature mapping method and the fixed Top-N weighted Gaussiancomponent feature mapping method. It can further improve the recognition performanceof the audio event recognition system which is under the case of channel mismatch.For further improving the channel adaptive ability of the mismatch system, unfold aresearch on the model domain channel adaptive methods and put forward a channeladaptive method of audio event recognition by audio event model synthesis (AEMS).Firstly, use the training data to complete the channel model and original audio eventmodel training; secondly, make the channel decision and generate the audio event modelfor the extracted test data in its channel type; finally, make the audio event modelrecognition based on the fragment and length. Experimental results show that theproposed method can remove the channel information, resolve the problem of channelmismatch and improve the ability of resisting channel distortion of audio event recognition system under the influence of network coding difference successfully.In addition, a prototype system of audio event recognition based on a joint channeladaptive method which combines the feature domain and model domain method isdesigned and implemented. Experimental results show that the system can achieve anaverage fragment F of88.67%and time length F of85.20%, has a better channel adaptiveperformance compared with the single domain method.