Supervised acoustic topic model for unstructured audio information retrieval

We introduce a modified version of the acoustic topic model, which assumes an audio signal consists of latent acoustic topics and each topic can be interpreted as a distribution over acoustic words, for unstructured audio information retrieval applications. The proposed supervised acoustic topic model is based on supervised latent Dirichlet allocation (sLDA) while the conventional acoustic topic model is built upon latent Dirichlet allocation (LDA) which learns its parameters in an unsupervised manner. The experimental results with BBC Sound Effects Library indicate that the supervised acoustic model brings benefits in terms of classification accuracy by learning parameters with respect to corresponding categories of audio clips, i.e., semantic and onomatopoeic labels. Index Terms — audio information retrieval, acoustic topic model, unstructured audio, supervised LDA