Effectiveness of HMM-Based Retrieval on Large Databases

Abstract

We have investigated the performance of a hidden Markov model based QBH retrieval system on a large musical database. The database is synthetic, generated from statistics gleaned from our (smaller) database of musical excerpts from various genres. This paper reports the performance of several variations of our retrieval system against different types of synthetic queries on the large database, where we can control the errors injected into the queries. We note several trends, among the most interesting is that as queries get longer (i.e., more notes) the retrieval performance improves.