We exploited deep neural networks (DNN) for two-to-five channel surround decoding. Specifically DNNs are used to replace the primary-ambient separation and ambient-signal-rendering modules. For the training, the mean-squared error of the magnitude spectra between the decoded and five-channel target signals and the interchannel level differences between the target signals were used as the loss functions. Through this procedure the DNNs can derive the spectral weights that can be used to produce the decoded signals, similar to that for the target signals. The log spectral distance, signal-to-distortion ratio, and multiple stimuli with hidden reference and anchor tests were used for objective and subjective evaluations. The experimental results show that exploiting the DNNs can generate decoded signals that are more similar to the target signals than those obtained via previous methods.
Choi, Jeonghwan; Chang, Joon-Hyuk
Affiliation: Hanyang University, Seoul, Republic of Korea
JAES Volume 68 Issue 12 pp. 938-949; December 2020
Publication Date: January 11, 2021
No AES members have commented on this paper yet.
If you are not yet an AES member and have something important to say about this paper then we urge you to join the AES today and make your voice heard. You can join online today by clicking here.