Transcribing Human Piano Performances into Music Notation

Automatic music transcription aims to transcribe musical performances into music notation. However, existing transcription systems that have been described in research papers typically focus on multi-F0 estimation from audio and only output notes in absolute terms, showing frequency and absolute time (a piano-roll representation), but not in musical terms, with spelling distinctions (e.g., A[ versus G]) and quantized meter. To complete the transcription process, one would need to convert the piano-roll representation into a properly formatted and musically meaningful musical score. This process is non-trivial and largely unresearched. In this paper we present a system that generates music notation output from human-recorded MIDI performances of piano music. We show that the correct estimation of the meter, harmony and streams in a piano performance provides a solid foundation to produce a properly formatted score. In a blind evaluation by professional music theorists, the proposed method outperforms two commercial programs and an open source program in terms of pitch notation and rhythmic notation, and ties for the top in terms of overall voicing and staff placement.