A tree-based method for fast melodic retrieval

The evolution of aurally queryable melodic databases (so-called query-by-humming systems) has reached a point where retrieval accuracy is relatively high, even at large database sizes. With this accuracy has come a decrease in retrieval speed as methods have become more sophisticated and computationally expensive. In this paper, we turn our attention to heuristically culling songs from our database that are unlikely given a sung query, in hopes that we can increase speed by reducing the number of matching computations necessary to reach the proper target song.