利用核依賴估計來進行多軌自動混音

Automatic Multi-track Mixing by Kernel Dependency Estimation

近年來由於數位音樂的蓬勃發展，錄音器材越來越普及。使得非混音專業人士也能利用錄音界面（Audio Interface）錄製出不錯的成品;但是一旦錄製了多軌（Multi-Track Recording）就會面臨到混音（Mixing）的問題，即需要把多軌的聲音混合在同一個軌中。混音牽扯到許多音響及聲學心理學的相關技術與知識，非專業人士要混出尚可的成品有一定的難度，所以我們提出了自動多軌混音系統（Automatic Multi-track Mixing System），希望藉由監督式學習的方式學習各軌間混音參數的調配，產生每首的基礎混音（Basic mix-down）來幫助非混音專業人士也能混出不錯的成品（Mix-down）。由於混音參數取得不易，我們會先藉由分軌及混音好的關係估計出各個混音參數，接著利用其參數進行混音模型（Model）的建立。在參數學習（Parameter Learning）方面由於每軌的混音參數是有依賴關係的（Dependency），我們採用了核依賴估計（Kernel Dependency Estimation）的參數學習（Parameter Learning）方式來預測每軌的混音參數。

Due to the revolution of digital music, people can create recordings in a home studio with cheaper gear. However multi-track recordings need to be mixed to combine them into one or more channels. The question is that mixing requires background knowledge in sound engineering and psychoacoustics. It is difficult to get good mixdown for non-specialist in sound engineer. In this paper, we use supervised learning method for automatically mixing multi-track recording into coherent and well-balanced piece. Due to lack of mixing parameters, first we estimate the weight of mixing parameters by using the relation between raw multi-track and mixdown. Given the mixing parameters for any music genre, we use kernel decency estimation method to create our mixing model. The experiment show KDE is able to make a more satisfactory estimation than treating each parameter independently.