Support vector machine-based automatic music transcription for transcribing polyphonic music into MusicXML

Abstract

Automatic Music Transcription (AMT) which transcribes music into music sheet is a challenging task since it requires combination of three different knowledges: signal processing, machine learning, and musical model. The task is more challenging when AMT applied to the polyphonic music. Such task required the system to recognize the pitch, timbre, tempo, onset, and expression into a readable music sheet. This paper describes our works in building such system. In this research, the most promising and prominent approach is applied. Those are the Mel's Frequency Cepstral Coefficient (MFCC) as the features and the One-against-all Support Vector Machine (SVM) as its decoder. The combination of both methods had shown very promising results. The output of our AMT system is a music sheet in a MusicXML format with high compatibility with music software nowadays.