自動鋼琴譜辨識

Abstract

樂譜辨識是一自動文件處理問題，將樂譜轉換成編碼形式後，可進一步發 展應用，例如樂曲風格分析、自舊版本重新編排成新版本、與音樂演奏等 。我們考慮的鋼琴譜辨識問題目的在於辨識出表示旋律與和聲的樂符但不 考慮表示如何演奏的樂符。不同意義的音符形態變異很大，各樂符間存在 著樂理關係，音符的音高決定於其所在的譜線垂直位置，但譜線的存在造 成許多樂符相連，增加樂符辨識的困難度。我們引用樂理知識，以小節為 處理單位，由小節線與譜線的關係將譜線定位並去除；根據樂符的標記法 則將樂符的辨識處理步驟分成兩部份：基本樂符元素抽取，樂符合成與辨 識。對於初學者演奏的鋼琴譜，本辨識系統的辨識率可達90%以上。 Music score recognition is a problem of document image analysis. After a music score has been translated to a music code, it can be used for further applications, such as analyzing musical style, re-engraving existing scores and producing audio version. The objective of the present exercise is to recognize symbols that describe what to be played in printed piano music, but not to recognize symbols that direct how to be played. The shape of notes varies from simple to difficult, and the pitches of notes depend on the vertical coordinate system defined by staff lines. However, the staff lines connect music symbols which would be normally be disconnected. This complicates the recognition process. We introduce musical knowledge in the proposed system. Our processing procedure is based on bar unit. The staff lines are detected and deleted by the relation to bar lines. According to the structure of music notion, the recognition process of music symbols is divided into two parts: primitive extraction, and symbol synthesis and recognition. Experiments demonstrate that recognition accuracy of some piano music for beginners exceeds 90%.