Follow That Tune — Dynamic Time Warping refinement for Query by Humming

Dynamic Time Warping is a standard algorithm used for matching time series irrespective of local tempo variations. This type of variability is inherent to audio input data obtained directly from users and, as such, it occurs in the context of Query-by-Humming interface to multimedia databases. Apart from the time-alignment problem, most of the known melodymatching approaches are also affected by a second issue of aligning the pitch between the query submitted by a user and the template. The query is usually in a different key and it may be simply sung out of tune, which needs some additional, sometimes computationally expensive processing and may not guarantee the success e.g. in the presence of pitch trend or accidental key changes. The method of tune following, proposed in this paper, enables to solve the pitch alignment problem in an adaptive way inspired by the human ability of ignoring typical errors occurring in sung melodies. The experimental validation performed on the database containing 4431 queries and over 5000 templates confirmed the enhancement introduced by the proposed algorithm in terms of the global recognition rate.