iComper:使用隱馬可夫模型於偵測音樂暗示之互動鼓手

Abstract

iComper是一能理解演奏者即興時所傳達的意圖之互動爵士伴奏系統。在本論文中，我們提出一個基於隱馬可夫模型之音樂信號偵測機制，透過這個機制橋接伴奏者在聆聽音樂時，透過所學之音樂理論及聆聽感受得到的理解與即興者演奏時傳達出的意圖。我們將即興者演奏的旋律以音樂理論為基礎之特徵擷取方法取出資訊並對應至觀察現象之代號。偵測機制會使用從樂曲開始到現下小節的觀察現象序列給予最可能的狀態。透過本文提出的鼓節奏產生演算法，組合從資料庫中取出的打擊樂器樣本以及音量變化樣本得到打擊樂器之伴奏。

iComper (interactive Comper) is an interactive Jazz accompaniment system which understands various soloist's intention when he or she improvises. In this thesis, we propose an HMM-based musical sign detector to bridge accompanist's realization when listening to music and intention given by the improviser. We apply music theory to our feature extractors on improviser's melody and map features to observations. The detector gives the highest-probability state using the observation sequence from the start of a tune to the current measure. A re-drum algorithm is proposed to combine the chosen percussion patterns and the volume pattern to generate accompaniment.