End-to-end Stereo Audio Coding Using Deep Neural Networks
Wootaek Lım, Inseon Jang, Seungkwon Beack, Jongmo Sung, Tae‐Jin Lee · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
Recently, deep neural-network-based audio data compression has been widely studied. Although most studies in this area have focused on mono channel speech coding, research on stereo coding is required for ultimate audio coding. In this paper, we consider in case of the stereo channel and present a method of end-to-end stereo audio coding, which exploits the concept of a mid/side (M/S) stereo coding scheme in the latent space of neural network. The goal of this study is to create an end-to-end neural network to find the optimal transformation for stereo coding. Therefore, the latent space conceptually corresponds to M and S, but better feature transformation for stereo coding is learned through the network. Through the end-to-end learning, the proposed method provides a more efficient bit allocation result than the discrete coding of each stereo channel into mono. According to the results of an objective test, the proposed method shows a superior performance in comparison to a discrete stereo model and conventional HE-AAC. A subjective evaluation shows that, for a bitrate of 64 kbps, the proposed model provides a significantly better sound quality than the discrete stereo coding model, as well as a performance comparable to that of a conventional HE-AAC.