時變性的音樂情緒成份分析研究

Abstract

人們的情緒隨著聆聽音樂的過程被牽動，以往音樂情緒檢索系統多半提出單一情緒分類。本研究以時變性的情緒變化為基礎並嘗試結合各種心理學家提出的情緒模型，包含二維情緒模型和類別式情緒模型，進而提出以Content、Depression、Anxious、Exuberant四種情緒為基礎成分並隱含二維情緒模型資訊的即時性音樂成分分析系統。系統中採用Support Vector Machine當作分類的演算法，針對各種特徵以192首訓練歌曲建立兩兩情緒間的分類模型以此考慮二維情緒模型中的兩個維度單獨分析的情況。特徵值上採用音樂特徵和音訊特徵兩大類，共計11個特徵。並依其特性以不同的音框長度作分析。最後再以音樂情緒的問卷調查來比對程式實驗結果和實際聆聽者感受是否符合。

While listening to music, people's emotions are affected. The common music emotion recognition systems used to provide only one emotion classification. This study presented a time-varying music emotion analysis, and tried to integrate several psychological emotion models, including two-dimensional emotion space and category type. And further to make a time-varying analytic system, based on four basic compositions: Content, Depression, Anxious and Exuberant, which also contains information from two-dimensional emotion space. This system uses Support Vector Machine as classified algorithm. This system uses Support Vector Machine as classified algorithm. Training in a variety of features by 192 music clips as training data to build emotional classification model between each two, in order to inspect the situations for analyzing the two dimensions separately. Feature extractions were divided into two categories: Music Features and Audio Features, total of 11 features. Each feature used different length of frame for analysis. In the end, this study performed a questionnaire survey to compare the program results with the actual listeners’ experiences.