Probabilistic Strategies Based on Staged LSH for Speedup of Audio Fingerprint Searching with Ten Million Scale Database
Masahiro Fukuda, Yasushi Inoguchi · 2017
We are developing and improving algorithms to identify audio fingerprints (AFP) in a network router. Staged Locality Sensitive Hashing (LSH) is one of them and nearly as fast as 1Gbps of prevalent network routers. In this paper, we propose two extensions from Staged LSH, both of which take advantage of probabilistic strategies. One is Neighbor Staged LSH, which is to tune up about how to choose buckets for the hash method for searching. The other is Hierarchy Staged LSH, whose strategy is to focus on the popularity of songs. Adopting both achieved at most 182.8 times as fast as the simple Staged LSH and it was equivalent to 1 Gbps. The accuracy rate was 100 % if the BER of AFP is less than 15 %.