Transcription of Sung Song

Abstract

Several efforts have been aimed at automatic transcription of instrument tone. This paper deals with a transcription method of sung song using digital signal processing techniques. The fundamental function of transcription is the measurement of tonality (key signature) and metrics (time signature). The tonality measurement is based on peak frequency identification and the metrical feature is obtained from segmentatoin of the vocal sound. The peak frequency is identified by the interpolation of its complex spectrum, which has been recently proposed by the authors. Unstable peak frequency is reduced by median filtering and pitch drift is compensated by histogram techniques. The segmentation is carried out by power dip, key change, and variation of harmonics ratio. The key scale is determined by comparing the tonality distribution with every major scale and minor scale. The metric unit is given by correlating the segmentation result with the standard rhythm.
Experimental resuits show that two Japanese nursery rhymes have been stably transcribed into acceptable notes. This system can be served as a supporting facility for the ethnomusicology.