Automatic Transcription of Singing Voice Signals

Abstract

Automatic music transcription refers to an automatic extraction of musical attributes such as notes from an audio signal to a symbolic level. The symbolized music data are applicable for various purposes such as music education and production by providing higher-level information to both consumers and creators. Although the singing voice is the easiest one to listen and play among various music signals, traditional transcription methods for musical instruments are not suitable due to the acoustic complexity in the human voice. The main goal of this thesis is to develop a fully-automatic singing transcription system that exceeds existing methods. We first take a look at some typical approaches for pitch tracking and onset detection, which are two fundamental tasks of music transcription, and then propose several methods for each task. In terms of pitch tracking, we examine the effect of data sampling on the performance of periodicity analysis of music signals. For onset detection, the local homogeneity in the harmonic structure is exploited through the cepstral analysis and unsupervised classification. The final transcription system includes feature extraction and probabilistic model of the harmonic structure, and note transition based on the hidden Markov model. It achieved the best performance (an F-measure of 82%) in the note-level evaluation including the state-of-the-art systems.