音樂訊號處理在實務上有著許多的應用,這些應用包括了資料庫查詢系統、自動音樂訊號辨識以及可以作為音樂家的寫作工具。而特徵值在分析樂器訊號處理上又是一個很重要的部分,這些訊號包含了許多的資訊,而特徵值的計算在訊號處理上是獲得特定重要資訊的一個步驟。 在本篇論文裡我們實作了一個樂器音色分類的模型。我們介紹了許多音樂訊號上特徵值的特性以及計算的方法,當然也包含了在頻譜分析下的特徵值。我們使用K-Nearest Neighbors 演算法作為我們樂器分類的基礎,作為測試的音樂檔案包含了絃樂器以及管樂器,我們使用決策樹的方法找出利用最少的特徵值來分辨出各種的樂器種類,並且達到一個完美的準確率。
Music content analysis usually has many practical applications. For example, such applications include database retrieval systems, automatic music signal annotation, and musicians’ tools. In this thesis, we present a system for musical instrument classification. A wide set of features covering both spectral and temporal properties are investigated and their extraction algorithms are designed. We apply the K-Nearest Neighbors algorithm as the classification method. The instrument samples included string (bowed and struck), woodwind (single, double, and air reed). Using the complete feature for training, we achieve a perfect accuracy. We apply decision trees to select the best feature subset to improve the identification performance.