Research On Indoor Sound Source Localization Based On Pattern Recognition

Pattern recognition has been widely used in sound source localization in the past few years.Although this kind of method has stronger robustness,but some problems such as large calculation and low positioning accuracy in the high-noisy and reverberant environment still exist.To solve these problems,an improved sound source localization method based on feature length has been proposed.By extracting the feature of generalized cross coerration phase transform(GCC-PHAT)function in different length,the locating performance of classification recognition based on naive bayes,linear discriminant analysis,and support vector machine was improved.The work of the article includes mainly:1.Selecting the appropriate GCC-PHAT length not only maintain a considerable positioning accuracy by using shorter GCC-PHAT,but also reduce the computation in a weak reverberant environment and the positioning accuracy could be improved by increasing the GCC-PHAT length in a strong reverberant environment.Experiment shows that the linear discriminant analysis classifier is better than the other two in a strong reverberant environment.2.Sound source localization method based on recognition can be optimized by selecting the appropriate GCC-PHAT length for sound source at 10,50,90 degree azimuth.Experiment shows that considerable positioning accuracy can be achieved by using shorter GCC-PHAT in a reverberant environment for sound source at 90 degree azimuth and higher positioning accuracy can be achieved by using longer GCC-PHAT in a reverberant environment for sound source at 10 degree azimuth.3.One with the highest positioning accuracy of the linear discriminant analysis classifiers trained by GCC-PHAT in different SNR was selected for indoor sound source localization.Experiment shows that considerable positioning accuracy can be achieved by using a linear discriminant classifier with similar SNR in a weak reverberant environment and considerable positioning accuracy can be achieved by using a linear discriminant classifier with equal or slightly lower SNR in a strong reverberant environment.