Understanding Search Performance in Query-by-Humming Systems.

Previous work in Query-by-Humming systems has left open many questions. Although a variety of techniques have been explored, there has been relatively little work to compare them under controlled conditions, especially with “real” audio queries from human subjects. Previous work comparing note-interval matching, melodic contour matching, and HMM-based matching is extended with comparisons to the Phillips CubyHum algorithm and various n-gram search algorithms. We also explore the sensitivity of note-interval dynamic programming searches to different parameters and consider two-stage searches combining a fast n-gram search with a more precise but slower dynamic programming algorithm.