運用深度類神經網路模型進行歌聲分離及音高抽取

Abstract

近年來類神經網路隨著平行處理的技術以及硬體快速的發展而再度盛行，越來越多的音訊處理採用類神經網路的優勢，而本篇論文使用類神經網路將聲音從背景音樂中分離，區分為時域及複數域，另外因深度類神經網路估測出的結果仍有改善空間，我們遂提出新的聲音復原後處理方法，來增強GSIR的分數以及提升總音高的正確率。

Recently, neural networks prevails again due to the progress of fast hardware with parallel processing capability. More and more audio processing tasks take advantage of neural networks to achieve better performance. This paper utilizes deep neural networks (DNNs) to separate singing voice from background music in real and complex domain, respectively. Because the output of DNN are still not good enough, we propose two new methods of voice recovery for improving the GSIR of vocal separation. The proposed methods also achieves better accuracy for vocal pitch extraction.