A Quantitative Evaluation of a Two Stage Retrieval Approach for a Melodic Query by Example System

We present a two-stage approach for retrieval in a melodic Query by Example system inspired by the BLAST algorithm used in bioinformatics for DNA matching. The first stage involves an indexing method using n-grams and reduces the number of targets to consider in the second stage. In the second stage we use a matching algorithm based on local alignment with modified cost functions which take into account musical considerations. We evaluate our system using queries made by real users utilising both short-term and long-term memory, and present a detailed study of the system’s parameters and how they affect retrieval performance and efficiency. We show that whilst similar approaches were shown to be unsuccessful for Query by Humming (where singing and transcription errors result in queries with higher error rates), in the case of our system the approach is successful in reducing the database size without decreasing retrieval performance.