High performance indexing for massive audio fingerprint data

Real-time online audio searching systems require high performance indexing architecture for massive audio fingerprints. To improve the performance of hash table based indexing for audio fingerprints, this paper designs and evaluates a hybrid data structure which combines linked list with vector to store the values in the hash table to balance the searching performance and the memory usage. To extend the hash table to cluster environment, three distribution patterns are designed and implemented, and experiments show that the content-oriented distribution pattern is better than the keyoriented distribution pattern. The proposed serialized data layout of the hash table can further improve the searching performance with less memory usage. All the experiments are executed on practical massive data sets including up to 1,000,000 songs, and the results certificate the improvement of the methods proposed.