206 Efficient Audio Retrieval Based on Locality Sensitive Hash

This paper proposes an audio retrieval method based on locality sensitive hashing (LSH). In audio processes, there are two important parts: for audio database how to construct an index, for queries how to find out the matching ones. The method constructs an index of audio fragments by extracting fingerprint vectors from a database of audio. In retrieval, for each query extracts FP vectors and divides feature vectors into blocks with constant size, for each query block the method searches for similar FP blocks in the database to obtain a list of candidate index information, and this is performed efficiently by using LSH. We arrange a matrix contain all lists based on blocks' sequence, and use the Sub-linear Dynamic Programming method to find out best matching result and second similarly audio. Based on the robust fingerprint feature and efficient search algorithm, our experimental results show that the retrieval accuracy can achieve 90% with a cost of few seconds.