使用GMM轉換之背景伴奏消除及趨勢估計之歌曲音高軌跡追蹤

Abstract

在音樂資料檢索中，要對音樂做任何的分類、搜尋或分析，都要先對資料庫中的音樂提取適當的描述作為比對基準，在描述一段音樂的多種基準中，音高軌跡是一種直觀有用的資訊，在有人聲歌唱的音樂中，它指的就是人類歌聲的音高變化，如何從多個音源的音樂中擷取出人類的聲音，並做自動的音高追蹤是本論文的研究重點。 本研究先使用在語音處理中常用來做語者聲音轉換的GMM轉換方程式，來去除音樂中的背景伴奏，將含有伴奏的特徵參數轉換成純人聲的特徵參數；接著進行人聲音高趨勢估計，目的在於預測出音高軌跡可能存在的範圍，用以縮小基頻搜尋範圍，除加快運算速度外，並可排除諧波對音高軌跡追蹤所造成的不良引響；最後以動態規劃或是直接峰值選取的方式完成音高軌跡追蹤。實驗結果顯示本研究所提出的方法和現有最好的方法成效相當。

In music information retrieval, a proper representation of music signals in the database is needed for music analysis, search or classification. Among all existing music representations, pitch contour is an intuitive and useful feature. In a song, pitch contour is the fundamental frequency variation of the singing voice. The main concern of this thesis is hence to automatically extract the pitch contour of singing voice from a polyphonic music. In this study, we first employ the GMM transform function, which is popularly used in speaker voice conversion, to remove the background accompanies of polyphonic music. It transforms the characteristic features of a polyphonic music to those of vocal-only signal. Then, a trend estimation is conducted to predict the range of human’s pitch contour from the vocal-only characteristic features. The trend estimate is used to restrict the searching region for speeding up the searching process as well as for eliminating the interference of harmonics on pitch tracking. Lastly, pitch contour is obtained by dynamic programming or simple peak picking. Experimental results confirmed that the performance of the proposed method was comparable to the best method existing nowadays.