Robust direction-of-arrival estimation for a target speaker based on multi-task U-net based direct-path dominance...

Area of Science:

• Signal processing
• Acoustic signal analysis
• Machine learning for audio

Background:

• Direction-of-arrival (DOA) estimation is crucial for audio source localization.
• Existing methods, like multi-task networks for direct-path dominance, show promise but require refinement.
• Two-stage processing can introduce inefficiencies in DOA estimation pipelines.

Purpose of the Study:

• To refine a multi-task network for direct-path dominance testing.
• To eliminate the need for two-stage processing in DOA estimation.
• To validate the improved robustness and generalization of the refined network.

Main Methods:

• Development of a single-stage, dedicated multi-task network.
• Comparative analysis against a single-task network using an expanded dataset.
• Validation in high-noise environments and real-world applications.

Main Results:

• The refined multi-task network significantly enhances the robustness of DOA estimation.
• The single-stage approach demonstrates superior performance and generalization capabilities.
• The method proves effective even in challenging, high-noise acoustic conditions.

Conclusions:

• The proposed single-stage multi-task network offers a more efficient and robust solution for DOA estimation.
• The refined network generalizes well across a larger dataset and diverse noisy environments.
• This advancement holds significant potential for real-world audio processing applications.