自動化歌曲分段和主副歌辨識

Abstract

自動化歌曲分析、創作等議題，ㄧ直以來都是音樂科技領域裡非常重要的主題。本研究提出在數位化格式資料中，透過主旋律萃取、音類集合理論等概念做為前置處理程序，接著將解決最長共同子序列問題的動態規劃演算法，套用在自動化歌曲分段，來達成自動化分段的目的。 此外，本研究也分析了歌曲結構中段落之間的轉換關係，利用各段落之間的平均轉換率當成參考依據，自動辨識歌曲的主歌和副歌，並且探討各種音樂風格與自動化分段和主副歌辨識之間的關係。本研究所有的實驗和分析歌曲，皆是隨機採用ISMIR在2011年所發表的百萬首歌曲資料組。

Automated song analysis and composition is one of the most significant issues in the field of music technology. In this research, we bring up the methods including melody extraction and pitch class set theory in pre-processing stage and apply dynamic programming which used to solve longest common subsequence (LCS) problem to segment the song through symbolic data. In addition, we also analyze the song structure, and calculate average transition rate of every state transitions as a reference of verse-chorus identification. And we probe the relationship among music genres, segmentation and verse-chorus identification. The whole songs of experiment and analysis in this research are selected randomly in million song dataset (MSD) published in ISMIR 2011 as a dataset.