Acoustic stopwords for unstructured audio information retrieval

The notion of acoustic stopwords is proposed to improve the performance of generic unstructured audio information retrieval systems. The rationale behind this is based on the assumption that not all portions of a generic audio signal contribute toward deriving specific descriptive categories (semantic words and onomatopoeias in this work). Detecting the non-salient regions in the audio can hence lead to more robust mapping of signal to categorical descriptions. Using the latent perceptual indexing (LPI) based framework, we propose to remove the proposed acoustic stopwords from the extracted acoustic features, which may include little information on descriptive categories. The acoustic stopwords are selected based on data-driven frequency-related measurements such as document frequency (DF) and inverse document frequency (IDF). The empirical results with BBC sound effect library show that removing acoustic stopwords based on the IDF measurement improves the audio classification performance especially for onomatopoeic labels.