The importance of F0 tracking in query-by-singing-humming

Abstract

In this paper, we present a comparative study of several state-of-the-art F0 trackers applied to the context of query-by-singing-humming (QBSH). This study has been carried out using the well known, freely available, MIR-QBSH dataset in different conditions of added pub-style noise and smartphone-style distortion. For audio-to-MIDI melodic matching, we have used two state-of-the-art systems and a simple, easily reproducible baseline method. For the evaluation, we measured the QBSH performance for 189 different combinations of F0 tracker, noise/distortion conditions and matcher. Additionally, the overall accuracy of the F0 transcriptions (as defined in MIREX) was also measured. In the results, we found that F0 tracking overall accuracy correlates with QBSH performance, but it does not totally measure the suitability of a pitch vector for QBSH. In addition, we also found clear differences in robustness to F0 transcription errors between different matchers.