Automatic Audio Upmixing Based on Source Separation and Ambient Extraction Algorithms
Marian Negru, Bogdan Moroșanu, Ana Neacşu, Dragoş Drăghicescu, Cristian Negrescu · 2023
The stereo format is currently the most used audio format, with the drawback that it is difficult to fully provide an immersive surround sensation to the listener. Furthermore, most of the old recordings can benefit from a multi-channel transformation that can obtain an improved spatial sound. In this paper, an upmixing method based on two modules is proposed, which converts an original 2-channel audio to a 5.1 format. We used a Deep Neural Network model that can separate 4 sound sources (drums, bass, vocals, other), as well as a Primary-Ambient extraction algorithm, in order to obtain the 6-channel audio. We considered the 4 sources as Primary sounds and used only the Ambient effects from the second module. This new method offers good results for both music and film recordings, making it independent to the audio type. Our algorithm is already implemented in an industrial application, and we are constantly updating it for a better performance.