2015-08-012024-05-17https://scholars.lib.ntu.edu.tw/handle/123456789/686607摘要：隨著音樂資料庫日益擴大，如何幫助使用者快速取得喜歡的音樂成為了重 要的研究課題。由於音樂的喜好是因人而異的，能夠針對不同使用者量身打造 音樂檢索將可以大大的提升使用者體驗。本計畫將結合不同影響音樂喜好的因 子，發展個人化的音樂推薦或者音樂檢索。 影響一個人是否喜歡一首歌曲的因素是非常微妙且龐雜的，即使已有不少 人員投入相關研究，目前也都未有定見。但是音樂喜好因子大致可以分成兩種 類型，一種是與音樂內容有直接相關的，例如歌手的音色或者音樂所傳達的情 緒等，我們計畫透過機器學習的方法進行自動辨識。另一種則與使用者當下情 境有關，例如使用者當下的心情，我們將利用生理信號探測與分析的技巧取得。 掌握分析音樂喜好因子的方法之後，我們可以藉由分析一個使用者喜歡的 歌與他當下的狀態，推論出他目前應該會喜歡何種類型的音樂，進而找尋資料 庫中條件相符的音樂提供給使用者，達成個人化音樂檢索的目的。 本計畫將分三年執行，主要的內容包括(1)音樂內容相關的音樂喜好因子擷 取，包含旋律辨識、歌手音色辨識、個人化之音樂情緒辨識以及基於音樂情緒 之高階音樂描述；(2)使用者情境相關的音樂喜好因子擷取，包含生理信號的量 測與分析，還有其與情緒關聯性的研究；(3)結合各種音樂喜好因子，建立個人 化的音樂檢索系統。<br> Abstract: As the amount of music content continues to grow rapidly, how to efficiently find favorite songs for users has drawn much attention in recent years. Since music taste differs from one person to another, tailoring music retrieval to each user is necessary for better user experience. The goal of this project is to develop personalized music retrieval by investigating various music taste factors. Modeling music taste or preference is a challenging task because music taste can be affected by many factors and in subtle ways. Although much effort has been devoted to it, no conclusive agreement has been reached. However, we find that music taste factors can be grouped into two categories, one is music content–based and another is user context-based. Examples of the former are low level acoustic features such as MFCC and melody or high level features like singing voice timber and music emotion. Since all these factors are tied with music signals, we plan to extract these factors by signal processing and machine learning. On the other hand, user context-based factors, such as user mood, are only available from the user at the moment of music listening. Therefore, we plan to utilize physiological signal processing to obtain such information. Once music taste factors are available, we can model the music preference of a user by analyzing the favorite songs and the physiological signals of the user. Then, personalized music is recommended to the user by retrieving songs that match the music preference of the user. This three-year project includes three major parts: (a) retrieval of music content-based factors, including melody extraction, singing voice timber recognition, personalized music emotion recognition, and high level music description based on music emotion contour; (b) user context-based factor retrieval, including the processing of physiological signal and the investigation of its relation to user mood; (c) system development, including building a personalized music retrieval system that takes multiple music taste factors into consideration.機器學習音樂資訊擷取音樂旋律辨識歌手音色辨識音樂情緒 辨識生理信號推薦演算法。Machine learningmusic information retrievalmelody extractionsinging voice timber recognitionmusic emotion recognitionphysiological signalrecommendation algorithm.