Whistle for music: using melody transcription and approximate string matching for content-based query over a MIDI database

Abstract

In this paper, we present a "Whistle for Music" system which enables users to retrieve MIDI format music by whistling a melodic fragment. Three essential components are query processing, MIDI preprocessing and an approximate search engine. For query processing, we have achieved a real-time and robust whistle-to-MIDI converter. For feature extraction, the proposed MIDI preprocessing can extract individual, local and global melodic descriptions from MIDI files. In order to match query with target, we extend an existing search engine into a fast approximate melodic matching engine. Based on the integration of those three components, the system can return a list of MIDI files that are ranked by how closely they match the whistling. The systematic evaluation for the query-by-whistling system is finally performed. The results show that careful measurement and objective comparisons can lead us to know the scaling trend about query and target. One encouraging aspect is that the performance can be predicted based on the evaluation methods.