基於MPEG-7的歌者辨識與歌唱評分系統

Abstract

有別於傳統的關鍵字搜尋模式，利用音樂片段查詢音樂資料庫的內涵式資訊檢索技術已成為目前多媒體資料檢索的趨勢。本論文的首要任務是透過國際影音標準MPEG-7的音訊描述元辨識未知歌聲片段在一音樂資料庫中對應的演唱者身份。主要是針對音訊頻譜包絡描述元進行降維度處理取得其音訊頻譜投影，再加入輔助性的音訊頻譜質心，以提升歌者辨識的正確率。至於歌唱評分機制，本論文利用MPEG-7的音訊頻譜包絡描述元，進行MIDI主旋律音高與受評歌聲的色度特徵比對。為了避免男女歌者起始音高不同的情形影響到評分的準確性，我們執行半音刻度轉換與色度特徵的摺疊對照，再透過動態時間伸縮將不同時間長度的歌聲進行比對以及量化評分。

Unlike the conventional keyword-based searching mechanisms which rely heavily on the correctness of the given text, the content-based music information retrieval techniques that use only a small segment of music signals has gained popularity nowadays for their efficiency and accuracy. The main goal of this study is to identify whether a segment of unknown sound is from a specific singer in the database by using the MPEG-7 audio descriptor. The proposed scheme applies dimension reduction techniques on the MPEG-7 Audio Spectrum Envelope, thus obtain the corresponding Audio Spectrum Projection. Also utilized is the Audio Spectrum Centroid which improves the identification accuracy. Another issue addressed in this study is the singing evaluation, where the grading of each pieces of solo singing is carried out by changing Audio Spectrum Envelope into Chroma features and compare with MIDI melody pitch. MIDI-tone and Chroma feature scale conversion are conducted so as to compensate for the initial pitch difference of males and females. Moreover, Dynamic Time Warping is exploited to account for the differences in lengths.