Supervised acoustic topic model with a consequent classifier for unstructured audio classification

In the problem of classifying unstructured audio signals, we have reported promising results using acoustic topic models assuming that an audio signal consists of latent acoustic topics [1, 2]. In this paper, we introduce a two-step method that consists of performing supervised acoustic topic modeling on audio features followed by a classification process. Experimental results in classifying audio signals with respect to onomatopoeias and semantic labels using the BBC Sound Effects library show that the proposed method can improve the classification accuracy relatively 10~14% against the baseline supervised acoustic topic model. We also show that the proposed method is compatible with different labels so that the topic models can be trained with one set of labels and used to classify another set of labels.