Improving Target Sound Extraction with Similarity-Aware Beamforming

In this article, the authors propose a new beamforming method called SIBF (Spectral Incremental Blind Feature) that improves upon the traditional MMSE (Minimum Mean-Squared Error) blind filtering algorithm in terms of extraction accuracy. The SIBF method leverages the independence of sources to create a more accurate representation of the target source, while also reducing the computational complexity of the algorithm.
To understand how SIBF works, it’s helpful to think of the problem like a game of mixing different frequencies together in a cocktail. Just as you might want to separate the different flavors in a cocktail, the SIBF method separates the different frequencies in the signal to improve its accuracy.
The authors explain that traditional blind filtering methods, such as MMSE, use a cascade of two algorithms: one for filter estimation and another for scaling. However, these methods can lead to errors in the extraction of the target source, especially when the sources are highly correlated. SIBF addresses this issue by introducing an online algorithm that uses a reference normalization process to improve the accuracy of the estimate.
The reference normalization process involves dividing the estimated sources by their corresponding references, which helps to reduce the impact of errors in the estimation process. This process is similar to adjusting the volume of different channels in a stereo system to create a more balanced sound.
To further improve the accuracy of the SIBF method, the authors propose using a source model that represents a joint probability between the target and reference sources. This allows the algorithm to make more informed decisions about how to separate the different frequencies in the signal.
The experimental results presented in the article show that the SIBF method outperforms the MMSE method in terms of extraction accuracy, especially in situations where the sources are highly correlated. Overall, the authors demonstrate that the SIBF method is a promising approach for improving the accuracy of blind filtering algorithms in signal separation applications.