Acoustic topic model for audio information retrieval

A new algorithm for content-based audio information retrieval is introduced in this work. Assuming that there exist hidden acoustic topics and each audio clip is a mixture of those acoustic topics, we proposed a topic model that learns a probability distribution over a set of hidden topics of a given audio clip in an unsupervised manner. We use the Latent Dirichlet Allocation (LDA) method for the topic model, and introduce the notion of acoustic words for supporting modeling within this framework. In audio description classification tasks using Support Vector Machine (SVM) on the BBC database, the proposed acoustic topic model shows promising results by outperforming the Latent Perceptual Indexing (LPI) method in classifying onomatopoeia descriptions and semantic descriptions.