Using the Fisher kernel method for Web audio classification

As the multimedia content of the Web increases techniques to automatically classify this content become more important. In this paper we present a system to classify audio les collected from the Web. The system classi es any audio le as belonging to one of three categories: `speech', `music' and `other'. To classify the audio les, we use the technique of Fisher kernels. The technique as proposed by Jaakkola assumes a probabilistic generative model for the data, in our case a Gaussian mixture model. Then a discriminative classi er uses the GMM as an intermediate step to produce appropriate feature vectors. Support Vector Machines are our choice of discriminative classi er. We present classi cation results on a collection of more than 173 hours of Web audio randomly collected. We believe our results represent one of the rst realistic studies of audio classi cation performance on \found" data. Our nal system yielded a classi cation rate of 81.8%.