即時性音樂情緒響應追蹤研究

Abstract

音樂為什麼那麼具有吸引力的原因在於它喚起我們的情緒感受，如何從音樂訊號中擷取一套有效的方法識別情緒為本篇研究的目標，本研究認為整首的音樂不會只造成聽者產生單一種穩定的情緒，而是隨著歌曲的行進過程中，有持續的情緒變化，應該是時間性的變化過程，因此本論文提出具時變性的音樂情緒追蹤系統，追蹤音樂內容所喚起聽者的情緒感受，並記綠不同情緒的時間長度比例，比例越高的情緒，聽者的感受也就越長久且強烈。本研究參考音樂心理學的文獻並引用Thayer提出的情緒模型，由200首已標記情緒的音樂片段之WAVE檔中萃取五種特徵，包括音量、音樂事件密集度、音色、調性與音程的和諧度，經過情緒記分以界定在情緒帄面上的分佈，用兩種分類器GMM、BPN進行訓練，並收錄一百首流行歌曲做為情緒辨識的測詴樣本，輸出採三種呈現方式，分別有靜態的情緒追蹤、時間比例與動態的情緒可視化界面，研究最後請66名大學生進行共五次的音樂情緒問卷調查做為客觀的結果驗證。

The reason for the general appeal of music lies in the emotional feedback that music offers to its listeners. So, how to extract an effective approach from music signal to identify emotions response is the main purpose. This work proposes that music arouse our emotion response turn to the next story with continuous playback, it should be a real-time alteration not a static state. For the reason, research proposes a sequential framework real-time system to tracking emotions evoked by musical signals and recording the rate of each mood total time length. The higher proportion of mood means listeners feeling are more strong and long-term. The research considers psychological factors and cites Thayer's mood model reference. Five feature sets are extracted from the 200 WAV file music clips labeled mood state and applying to「Emotion Scores Counting」to define the distribution of sample space. Next, training two classifier (GMM, BPN) and recognize emotion by 100 Pop music testing data. Finally, the system has three methods (static emotion tracking, time-lasted rate and dynamic emotion locus visualization) to exhibit our result. Beside the researchers invite 66 college students attending total of five times music mood questionnaire survey to statistics an objective result as research verification.