互動式爵士鋼琴自動伴奏系統

Abstract

本篇論文目標為實作一個互動式的爵士鋼琴伴奏系統ICYS，採用樣本學習來建立模型。第一步先將整個問題一分為二，節奏還有旋律和聲。為了後續處理的方便，找出可以定性定量描述的節奏與旋律和聲的方法就變成本篇論文的第一步，在此定義了節奏向量代表節奏的特徵，以及使用和弦向量來訓練和弦和聲模型。為了讓未來有實作成即時伴奏的可能性，在這兩個子問題中都使用同樣的策略，先行預測再對預測的結果做伴奏。在預測的步驟中，節奏與旋律和聲的資料型態不一樣，所以因地制宜採用類神經網路或可適性線性預測等等的方法，依使用者的習慣逐漸學習；對於預測結果的伴奏，則是使用資料庫比對出最適合的節奏，利用和弦模型調整和聲，並且整合這兩邊的參數成為最後的伴奏。

In this thesis, a system ICYS (Interactively Comp Your Solo) that can accompany a jazz soloist as a jazz pianist is proposed. The learning scheme of the system is sample based. Under the concept of “divide and conquer”, the accompaniment procedure is divided into two independent paths, rhythmic and melody-harmonic accompanying. To accomplish real-time accompanying of the coming solo sequences, each of the two paths begins with analysis and prediction. Considering the different features of the data, several methods, like neural network and edit distance, are applied in predicting. After the prediction process, the next issue is to rhythmically and harmonically accompany according to the prediction results. At the end of this thesis, the audience feedback and evaluation of experts will be shown and discussed.