We have described a system for retrieving pieces of music from a database on basis of a sung query. The database is constructed automatically from a set of MIDI files, with no need for human intervention. Pieces in the database are represented as hidden Markov models (HMMs) whose states are note transitions. Queries are treated as observation sequences and pieces are ranked for relevance by the Forward algorithm. The use of note transitions as states and the Hidden Markov approach make for a system that is relatively robust in the face of key and tempo change. The use of observation probability distributions for hidden states deals with systematic error in query transcription. Hidden Markov models are an excellent tool for modeling music queries. The results of our experiments with this "first-step" implementation indicate both the promise of these techniques and the need for further refinement. Refinements to the hidden model topology and of the observation model will allow us to model a broader range of query behavior, and improve the performance of the system.
ASJC Scopus subject areas
- Library and Information Sciences