Implementation of a matching engine for a practical query-by-singing/humming system

This paper proposes a matching engine of a query-by-singing/humming (QbSH) system of which database is constructed from polyphonic recordings such as MP3 files. Use of the database makes the system more practical since it saves the trouble of gathering MIDI files. The pitch sequences transcribed from polyphonic recordings may have errors, and to reduce the influence of the errors, the matching engine uses chroma-scale representation, compensation, and asymmetric dynamic time warping. We propose the use of saturated distances, and it is verified that the distances perform better then generally-used absolute difference and squared difference. In our experiment, our QbSH system achieved mean reciprocal rank of 0.725 for 1000 singing/ humming queries when searching from a database of 28 hour audio data.