複音音樂的樂器編制分析及其在音樂相似度估計上的應用

Abstract

在過去的數十年裡，由於網際網路的蓬勃發展，各式各樣多媒體檔案的數量不斷增加。在這之中，不論是在獲取或是發佈數位音樂檔案都變得比過去容易很多。也由於此數量規模的不斷爆增，我們需要一個新的聆聽音樂和發掘新音樂的方法。 在這篇論文的一開始，我們會介紹一個簡單的音樂相似度估計系統，並且模擬它的效能。根據實驗結果顯示，使用較低階的特徵向量來描述曲子的特性，並不足以讓我們分離出不同音樂內容本身對相似度造成的影響，例如和絃、曲風、樂器編制和旋律。因此，這篇論文的主要目標在於將原本的低階特徵替換為與音樂內容有關的中階特徵。於此，我們將特別著重於樂器編制的自動化分析。音樂訊號音色的時頻分析和單一樂器的分類問題都將在此篇論文中討論，以做為基本的工具。之後我們將延伸此想法到處理更複雜的複音音樂，並且截取其隨時間變化的樂器編制資訊。藉由在相似度估計系統上使用此資訊，我們發現計算出的相似樂曲結果中，將可以特別針對樂器和音色，而非其他音樂內容。如此將可以取代原本的相似度估計系統，達到實現多模式音樂相似度估計的目標。

During the past few decades, the world has ushered in a new era, with booming Internet technology and immense multimedia content distribution. The acquisition and circulation of digital music ‾le become much easier than ever. Due to this rapidly rising of music quantity, a brand new way of discovering and recommending music is thus highly expected. In the beginning of this study, a conventional music similarity measure system based on the signal analysis methods is implemented and evaluated. According to the experimental results, it shows that the low-level features from signal analysis techniques are not strong enough to ful‾ll the discrimination between various musical content, such as the chord progression, genre, instrumentation, and melody. Therefore, the aim of this study is to incorporate the low-level feature with the mid-level feature, in order to utilize the musical content. We focus on the way to extract the instrumentation information leaved by the composers. The time-frequency analysis of musical instrumental signals and the classification problem of various instruments in the monophonic case are studied. After that, we extend the idea to deal with the polyphonic music and analyze its time-varying instrumentation information. By incorporating this information back to the original similarity measure system, the calculated similar songs can resemble to each other specifically in the sense of the instrumentation.