以自動簡化為基礎的音樂自動分析研究

Abstract

本論文首先探討各種自動作曲系統，從中歸納出對於自動作曲所需的重要音樂參數。接著利用電腦程式來進行自動分析音樂的工作，其中有些參數可直接利用統計方式得到，另外調性以及和聲的分析則是由各種相關文獻當中，找出特殊的演算法來達成。 在研究調性和和聲的分析時，我們發現了實驗中的誤差有一部份原因來自於樂曲中裝飾性的音符所產生。本文提出一個先將音樂自動簡化的演算法(AMSA, Automated Music Simplification Algorithm)來提升分析之準確度，文中比較了沒有簡化與簡化後之分析結果，在調性分析和和聲的分析上可增加約百分之六到百分之八的準確率。經由音樂自動簡化的手法，使得樂曲架構能更清楚顯現，也讓音樂參數分析能更加準確。 透過電腦自動化分析，可使需要大量人力與時間來進行的音樂分析減少成本。另外，可利用自動分析後的結果建立樂曲風格資料庫，提供各種風格的音樂規則來讓電腦進行自動作曲，達到能自動學習音樂風格並且自動創作的最終目標。

In this thesis, we investigate systems of automated composition and determine which music features are necessary for algorithm composition initially. Then this work build some automated tools to analyze music features such as pitch distribution, pitch transfer probability, key, and harmony. In the process of analyzing music, the accuracy will generally be affected by ornaments. A musical simplification algorithm AMSA (Automated Music Simplification Algorithm) was proposed to filter ornaments out in music. The comparison results were shown that the accuracies of key finding and chord finding have been improved about 6%~8% by using simplification method. After simplification, the music structure is getting more clearly than original one. Through the automatic music analysis, we can minimize the cost of music analysis. We can also use the analysis result to build a music style rule database. Eventually it is easily to compose music with style learning automatically using these rules.