Abstract: A novel method for audio feature extraction and recognition is presented In this method, FBM （fractional brownian motion） based fractal dimension is defined as audio fractal feature According to Gaussian distribution characteristic of audio fractal feature, Ada boosting algorithm is used for feature reduction Then two classifiers, weighted Ada Gaussian classifier and support vector machine, are implemented respectively for audio classification Based on these two classifiers, a multiple classifier model is finally constructed Experimental data shows that audio fractal feature achieves better performance than other audio features for music and speech classification