RELATIVE PITCH TRACKING FOR SINGING VOICE AS AN APPLICATION IN QUERY BY HUMMING SYSTEMS

摘要

Pitch extraction from singing voice has traditionally been viewed from the perspective of absolute pitch of the voice. However, for the query by humming applications, the absolute pitch is not of as much importance as compared to that of relative pitch. The algorithm presented in this paper addresses the issue of relative pitch tracking for singing voice and attempts to improve the transcription accuracy. It also tries to bring robustness to transcription, in the several forms of querying by a user. The algorithm makes use of specially designed Bach filters, which associate a relative frequency with a corresponding musical semitone. This simplifies the conversion of the transcribed pitch values to the MIDI format. Preliminary results show up to 85% accuracy in the automated transcription when compared with manual transcription.