Towards intelligent string matching in query-by-humming systems

In the last 5-10 years, there have been several attempts at the creation of a musical database that can be queried acoustically. The main problem hindering such efforts so far has been scalability, with limits of the frequently used three-character representation for melodic contour becoming apparent even at relatively small database sizes. Extending the contour representation to 24 semitone resolution would go a long way towards solving such problems, but applying current string matching methods to a 24 character alphabet would prove difficult at best, and human errors would be far more prevalent. Here we show that the vast majority of errors at this resolution, although more numerous, are quite predictable using current methods of probabilisic modeling. This is evidence which we hope opens the door to more intelligent forms of note matching.