頻譜長條圖、週期性長條圖、抖動模式在器樂相似度上之效果比較

Abstract

在這篇論文中，我將會詳細地介紹三種不一樣的作音樂相似度分析的方式。接著利用自己選出來的50首歌(粗略分為五個種類，每一個種類有十首歌)，依序用這三種方式來對它們做相似度的計算。比較計算出來的結果，我們可以得知哪一種方法最具正確性、最有效，可以將不同種類的音樂分得比較清楚。 因為一個音樂的檔案可以用多種不同的角度來看它，我們必須先決定要比較的是音樂在哪一種特徵上的相似度。在這篇論文中，我選擇「音色」作為我的分類標準。在第二章中，我會對音樂的「音色」這個特徵，做比較詳細的描述。其他的特徵諸如旋律、節奏等等也會稍微提到。 在第三章中，將會說明在語音辨識中被廣泛應用的「梅爾-倒頻譜係數」。另外一種類似「梅爾-倒頻譜係數」的方式--「宋(量測聲音大小的單位)-巴克表示法」也會被一起提到。他們是第四章所要說明的三種方法的基礎。 第四章提到三種比較音樂相似度的方式，也就是頻譜長條圖、週期性長條圖、以及抖動模式。第五章會把這三種方式應用到器樂上，藉以得知哪一種方式在分辨不同的樂器上有最好的效果。第六章則是結論，以及未來可以繼續進行研究的方向。

In this thesis, I introduce three different methods for music similarity measuring in detail. And then I apply these three methods to 50 songs I chose (frankly viewed as 5 groups, each one has 10 songs). By comparing these results, we can understand which method is the most effective and most accurate. Because we can view an audio data in many different views of point, we have to choose a property on which we focus before starting the similarity measures. In this thesis, I choose the property “timbre” as my yardstick. In chapter 2, I describe the property “timbre” in detail. Other properties such as melody, rhythm, and genre are referred to. Based on these properties, the very first step of measuring similarity of music is constructed. In chapter 3, some backgrounds of building these similarity measure methods are discussed in detail, e.g. Mel-Frequency Cepstrum Coefficient and Sone-Bark Representation. In chapter 4, three methods for measuring the music similarity are particularly described. They are spectrum histogram (SH), periodicity histogram (PH), and fluctuation pattern (FP). In Chapter 5, SH, PH, and FP will be applied to instrumental music. We can know the best method to distinguish different instruments. Chapter 6 is about conclusion and future works.