Foreground-Background Audio Separation using Spectral Peaks based Time-Frequency Masks

The separation of foreground and background sounds can serve as a useful preprocessing step when dealing with real-world audio signals. This work proposes a foreground-background audio separation (FBAS) algorithm that uses spectral peak information for generating time-frequency masks. The proposed algorithm can work without training, is relatively fast, and provides decent audio separation. As a specific use case, the proposed algorithm is used to extract clean foreground signals from noisy speech signals. The quality of foreground speech separated with FBAS is compared with the output of a state-of-the-art deep-learning-based speech enhancement system. Various subjective and objective evaluation measures are computed, which indicate that the proposed FBAS algorithm is effective.