Measuring Similarity of Automatically Extracted Melodic Pitch Contours for Audio-based Query by Humming of Polyphonic Music Collections

A study of melodic similarity of pitch contours automatically obtained from audio files in the context of Query by Humming is presented. Pitch contours are extracted directly from monophonic (query files) and polyphonic (commercial songs) audio files using a state-of-the-art algorithm MELODIA for automatic estimation of predominant melodic contours. The contours are then coded using the Symbolic Aggregate Approximation algorithm for the reduction of the huge amount of information each sequence contains, avoiding any step of automatic music transcription, and then compared using the subsequence matching and time warping method Smith-Waterman, being then an audio-based comparison. Results using the commented approach do not reach state-of-the-art results obtained by other authors in audio-based Query by Humming and, analyzing the results, the main conclusion is that Symbolic Aggregate Approximation might not be appropriate for this task