以圖像化音樂情緒分類系統應用 於音樂風格分析及曲目選取

Abstract

本研究建立一套以音樂訊號分析、二維情緒分布模型與機器學習分類器為基礎而發展出的圖像化情緒分類選歌系統，目的為將音樂多階層組合的序列傳遞訊息數據量化，發展使用者與手持裝置間之音樂訊息傳播媒介與方式，並希望藉此明確地建立不同使用者對音樂專輯的情緒認知與喜好分布。 實驗中先將音樂訊號取樣分段框化，擷取音訊特徵，計算此音訊特徵與音樂產生情緒的關聯，並以Thayer提出的:1.平靜的、2.興奮的、3.憂鬱的、4.生氣焦慮的四種情緒模型歸類，開發Android系統之圖像化音樂選取應用程式。本實驗選擇200首西洋流行通俗風格曲目(四類情緒各50首)作訓練資料，經情緒萃取演算法擷取特徵及輸入資料庫後，將計分情況繪出在情緒座標平面上，再以SVM(支持向量機)當作分類器訓練出情緒分界，輸入至手持裝置，除可顯示使用者音樂曲目聆聽喜好，亦可供使用者依據情緒喜好挑選行動裝置內的音樂。 實驗成果顯示在處理音樂數據上，以數理分析與擷取情緒特徵演算，並搭配辨識器分界訓練，開發出一套顯示聆聽者情緒回饋的人機互動系統。此研究之未來應用不只可出現在手機平板平台上，更能在車用配備、家用娛樂多媒體產品，提供創新音樂曲目選歌播放系統，並可依世界各地區之音樂曲風文化，建立客製化之音樂情緒座標平面，如中東音樂，宗教音樂等。

This paper proposes a sequential framework that progressively extracts the features of music and characterizes music-induced emotions in a predetermined emotion plane. To build-up the emotion plane, 200 Western pop music clips, including four categories of emotion-predefined music with each group of 50 clips are used to train the system. Five feature sets, including onset intensity, timbre, sound volume, mode, and dissonance are extracted from WAV file to represent the characteristics of a music sample. Support vector machine (SVM) algorithm is used to demarcate the boundaries of “Exuberance”, “Contentment”, “Anxious”, and “Depression” on the Thayer’s emotion plane for trained music data. A graphic interface of emotion arousal locus on two-dimensional model of mood is established in Android System to represent the tracking of dynamic emotional transition caused by music. This system enables user to choose playing clips in mobile device based on identifying music emotions. Furthermore, the graphic interface of emotion can also reveal the emotional distribution of user’s music clips in data bank. The experimental results show the exploiting of human-machine interaction system by the efforts of mathematical analysis, feature extraction and algorithm, classifier training in music data processing. This interactive music selection system provides innovative music tracks playing sequence. With various music genre of different regions in the world, such as the Middle East music, religious music, a customized music emotions coordinate plane based on specified training music samples can be created to show the diversity of music culture.