Automatically Identifying Vocal Expressions for Music Transcription

Music transcription has many uses ranging from music information retrieval to better education tools. An important component of automated transcription is the identification and labeling of different kinds of vocal expressions such as vibrato, glides, and riffs. In Indian Classical Music such expressions are particularly important since a raga is often established and identified by the correct use of these expressions. It is not only important to classify what the expression is, but also when it starts and ends in a vocal rendition. Some examples of such expressions that are key to Indian music are Meend (vocal glides) and Andolan (very slow vibrato). In this paper, we present an algorithm for the automatic transcription and expression identification of vocal renditions with specific application to North Indian Classical Music. Using expert human annotation as the ground truth, we evaluate this algorithm and compare it with two machinelearning approaches. Our results show that we correctly identify the expressions and transcribe vocal music with 85% accuracy. As a part of this effort, we have created a corpus of 35 voice recordings, of which 12 recordings are annotated by experts. The corpus is available for download 1 .