Unsupervised Learning Hash for Content-Based Audio Retrieval Using Deep Neural Networks

摘要

Binary hashing is an attractive approach for large-scale audio collection search, due to its encouraging efficiency in both speed and storage. However, most existing hashing methods for content-based audio retrieval assume that data are independently and identically distributed. In this paper, we propose a novel unsupervised learning hash method for audio retrieval. By utilizing the deep network for unsupervised learning audio representations, the compact binary codes can directly be generated at one layer. We also include similarity preserving property to enhance the quality of the binary hash codes and thereby increasing the overall accuracy of the audio retrieval. This method is evaluated on Ballroom datasets with variety of eight genres of 678 audio clips. The results of the experiment support the effectiveness for audio retrieval with high precision and recall values at 98.92% and 91.50% respectively.