Content-based Fast Audio Retrieval

As the rapid development of the applications on Internet, the traditional Text-Based Audio Retrieval Engine is losing its ability to satisfy user's experiences which are increasing every day. In recent years, the technology of Content-Based Audio Retrieval (CBAR) tends to be focused by scholars both at home and abroad. However, in face of a great deal of audio data on Internet, the technology of CBAR needs to be enhanced seriously both on the speed of retrieval and the robustness against noise. In this paper, the two problems above are focused, and related research of retrieval on audio examples at expression level is fulfilled.In terms of feature extraction of audio, which belongs in front-end part of the system, an audio retrieval method based on Local-Sensitive Hashing (LSH) index which has perfect theoretical framework and high retrieval speed is reviewed briefly. Aimed at the weakness of LSH index that a great deal of high-dimensional vectors should be compared in the corresponding bucket, a method of feature compression is proposed. Matched with proper back-end search schemes, the method proposed could avoid the process of comparing high-dimensional vectors in LSH-index method.In terms of back-end search schemes of audio features, the Inverted index which is widely used in the areas of natural language processing is reviewed briefly. Aimed at the weakness of basic Inverted index which could not reflect the time-series of audios, k-word proximity search is quoted, and an audio retrieval method based on Inverted index using compressed feature is proposed. Experimental results show that our scheme has higher speed than LSH-index method, and gets pretty fine precision when there're not so many noises in the test audios.Aimed at the problem that the robustness against noise of LSH-index method is not strong, an audio retrieval method based on gliding windows is proposed, which is optimized in terms of calculating speed later. Experimental results show that the robustness against noise of our scheme is better than LSH-index method, which has approximate speed of the latter.