Loudness normalization based on clean dialogue loudness improves consistency of the dialogue level compared to the loudness of the full program measured at speech or signal activity. Existing loudness metering methods can not estimate clean dialogue loudness from mixture signals comprising speech and background sounds, e.g. music, sound effects or environmental sounds. This paper proposes to train deep neural networks with input signals and target values obtained from isolated speech and backgrounds to estimate the clean dialogue loudness. Furthermore, the proposed method outputs estimates for loudness levels of background and mixture signal, and Voice Activity Detection. The presented evaluation reports a mean absolute error of 1.5 LU for momentary loudness, 0.5 LU for short-term and 0.27 LU for long-term loudness of the clean dialogue given the mixture signal.
(See document for exact affiliation information.)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
