Automatically extracting performance data from recordings of trained singers.

Abstract

ABSTRACT - Recorded music offers a wealth of information for studying practice. This paper examines the challenges of automatically extracting information from audio recordings of the singing voice and discusses our technique for automati catty extracting information such as note timings, intonation, vibrato rates, and dynamics. An experiment is also presented that focuses on the tuning of semitones in solo soprano performances of Schubert's Ave Maria by non-professional and professional singers. We found a small decrease in size of intervals with a leading tone function only in the non-professional group. KEYWORDS - vocal intonation, analysis, audio annotation This paper describes the challenges that arise when attempting to automatically extract pitchrelated data from recordings of the singing voice. The first section of the paper provides an overview of the history of analyzing recorded performances. The second section describes an algorithm for automatically extracting data from recordings of the singing voice where a score of the is available. The algorithm first identifies note onsets and offsets. Once the onsets and offsets have been determined, intonation, vibrato, and dynamic characteristics can be calculated for each note. The main experiment of the paper, described in the third section, is a study of intonation in solo vocal performance, where both the note onsets and offsets and fundamental frequency were estimated automatically. In the study, six non-professional and six professional sopranos performed Schubert's Ave Maria three times a cappella and three times with a recorded piano accompaniment. Our analysis of these recordings focused on the intonation of ascending and descending semitones. We found that the A-Bl, intervals with a leading tone function were on average 8 cents smaller than the non-leading tone A-Bb, but that their average size was approximately the same as the other semitones performed in the piece, regardless of intervallic direction or accompaniment. PREVIOUS WORK ON THE ANALYSIS OF RECORDED PERFORMANCES Interest in studying recorded performances dates back almost as far as the birth of recordable media, beginning with Dayton Miller's (1916) work on visualized pitch information in recordings with phonophotographic apparati. The psychologist Carl Seashore and colleagues at the University of Iowa also undertook extensive work in analysis (Seashore, 1938) employing a number of techniques to study recorded performances. Piano performances were studied from both piano rolls and films of the movement of the hammers during the performance. The team also undertook numerous studies of singing. Schoen (1922) studied five performances of Gounod's setting of the Ave Maria. He found that tuning depended on the direction of the line: notes following a lower tone tended to be flatter whereas notes followed by a higher note tended to be sharper. In general the singers were sharper than either equal temperament or just intonation. Easley's study of vibrato in opera singers found that the rate of the singer's vibrato was faster and the depth broader in songs from opera as compared to concert songs (Easley, 1932). Bartholomew (1934) studied vibrato along with other acoustic features of the singing voice in an attempt to define good singing. He observed the vibrato to be sinusoidal in nature and its rate to be approximately 6-7 Hz. H. G. Seashore (1936) also looked at Gounod's setting of the Ave Maria, as well as Handel's Messiah. He studied nine performances and focused on the connections, or glides, between notes. He was able to correlate glide extent with direction, finding that glide extent was larger going up than going down. Miller (1936) provided a large amount of detail through performance scores, though a lot of the data was not analyzed. His study of vibrato found that the rate of vibrato fell between 5. …