Kernel-based audio classification

Audio classification is subject to the heavy computation because of the high dimensionality of audio features as well as the unfixed length of audio segments. In this paper, an audio classification method based on the kernel is proposed, which could significantly reduce the dimensionality of audio features, and convert the variable length audio segments to fixed one. Gaussian Fisher Kernel is employed for transforming the audio clip to the equivalent parameter space, which bears the characteristic of low dimensionality. Audio reduct is extracted by the method of Variable Precision Rough Set Model, and it has the strong discrimination ability and could serve as the proxy of audio clip. Audio retrieval experiments show that our method could achieve the more accurate classification than conventional methods.