基於節拍同步和音色不變量之音色頻譜和交叉遞回圖分析之翻唱歌曲辨識系統

Abstract

辨認翻唱歌曲對人類來說是一件輕而易舉的事情。然而對電腦來說，有效率又準確的辨認翻唱歌曲並不是一件簡單的事情。一首歌曲可以用各種不同的方式來翻唱，例如：重新混音編曲，現場演奏，或是純樂器演奏版本等等，而這些版本和原唱版本之間又有各種不同的聲音特性的差異。本論文中討論了各種不同的Chromagram在翻唱歌曲辨識系統中的效能，並提出了一個基於節拍同步和音色不變量之音色頻譜和交叉遞回圖分析的翻唱歌曲辨識系統。這個系統在covers80 資料集的八十首歌中成功的辨識出六十二首歌。

Identifying cover version of a song is easy and straightforward for human. However, it still can not perform accurately by a computer. Every music recording can be covered in various ways, such as live performance, rearrangement, and instrumental...etc. Consequently, there are various musical variations between different cover versions and original version. According to the second hand song website, ”Yesterday” performed by ”The Beatles” has 233 covered versions. In this thesis, I attempt to explore the effectiveness of various chroma feature for cover song identification, and propose a modelfree system based on beat-synchronous time-invariant chromagram and cross recurrence plot analysis. Using the online available covers80 dataset, the numbers of correctly identification covers is 62 over 80 songs. The evaluation result also reveals that the enhancement of chroma feature will lead to dramatic performance gains.