- 學位論文
以深度學習實作流行音樂自動分類機制
Automatic Pop Music Classification Using Deep Learning
摘要
近年來科技持續在進步，現今社會面臨資訊量爆炸的問題，因此自動分類是不可或缺的角色，在音樂領域中也逐漸發展到高峰，要從巨量的音樂中挑選出自己喜歡的音樂歌曲，所花費的時間成本相對也會提升。一首音樂歌曲會受到喜愛，音樂特徵是個重要因素，每首音樂歌曲都會有獨特的音樂特徵，每種音樂類型的音樂特徵會是相似的，個體對相似的音樂特徵的評價應當會是一致的。人們往往會被演奏者的歌聲或是歌詞所吸引，導致有時音樂特徵通常會被忽略或是誤判。若能精準的分析每首音樂歌曲的音樂特徵，將與個體喜愛的音樂歌曲相似的的音樂特徵進行分類，相信是可以被接受甚至喜愛的。 本研究收集音樂公開資料集的音樂歌曲作為研究樣本，將每首音樂歌曲透過音檔格式資訊轉換為頻譜圖作為本研究的分類模型樣本，再透過卷積神經網路建立本研究的分類模型。收集華語音樂歌曲作為本研究的驗證系統樣本，將華語音樂歌曲透過卷積神經網路擷取的每種音樂歌曲類型的特徵資料進行分類，將與使用者喜愛的音樂歌曲相同的音樂類型與使用者進行互動得到回饋，透過回饋驗證本研究音樂歌曲自動分類的的準確率。
並列摘要
In recent years, technology has continued to advance, and today's society is facing the problem of information explosion. Therefore, the automatic classification is an indispensable role, and it has gradually developed to the peak in the music field. Have to select your favorite song from a huge amount of music, the cost of time spent will also increase. A music will be loved, music characteristics are an important factor, each song will have unique music characteristics, the music characteristics of each music type will be similar, and each people's evaluation of similar music characteristics should be consistent of. People are often attracted to the singer's singing or lyrics, resulting in music features that are sometimes ignored or misjudged. If we can accurately analyze the music characteristics of each song, and classify the musical characteristics similar to the music songs that individuals love, it can be accepted or even loved. This research collects songs from the public music dataset as research samples, and each song was converted into the spectrogram through the audio file format information as the classification model sample of this research, and then the classification model of this research was established through the convolutional neural network. Collect Chinese music songs as a sample of the verification system for this research, classify the feature data of each music song type extracted by the convolutional neural network of Chinese songs, and classify the music types and usages that are the same as the music songs that users love .The participants interacted to get feedback, and the accuracy of automatic classification of music in this study was verified through the feedback.