High-performance audio matching with features learned by convolutional deep belief network

Abstract

Audio matching automatically retrieves all excerpts that have the same content as the query audio clip from given audio recordings. The extracted feature is critical for audio matching and the Chroma Energy Normalized Statistics (CENS) feature is the state-of-the-arts. However, CENS might behave unsatisfactorily on some audio because it is a handcraft feature. In this paper, we propose to utilize the features learned by Convolutional Deep Belief Network (CDBN) to enhance the performance of audio matching. Benefit from the strong generalization ability of CDBN, our method works better than CENS based methods on most audio datasets. Since the features learned by CDBN are binary-valued, we can develop a more efficient audio matching algorithm by taking the advantage of this property. Experimental results on both TIMIT dataset and a simulated music dataset confirm effectiveness of the proposed CDBN based method comparing with the traditional CENS feature based algorithm.