[Paper Review] DOA Estimation by DNN-based Denoising and Dereverberation from Sound Intensity Vector

Masahiro Yasuda, Yuma Koizumi+3 more|arXiv (Cornell University)|Oct 10, 2019

Speech and Audio Processing7 references4 citations

TL;DR

This paper proposes a hybrid DOA estimation method that enhances sound intensity vector (IV)-based direction-of-arrival estimation by using two deep neural networks (DNNs) for denoising and dereverberation. The DNNs refine the IVs in the time-frequency domain before applying physics-based DOA estimation, achieving an average DOA error of 0.528°, significantly outperforming conventional IV-based and end-to-end DNN methods.

ABSTRACT

We propose a direction of arrival (DOA) estimation method that combines sound-intensity vector (IV)-based DOA estimation and DNN-based denoising and dereverberation. Since the accuracy of IV-based DOA estimation degrades due to environmental noise and reverberation, two DNNs are used to remove such effects from the observed IVs. DOA is then estimated from the refined IVs based on the physics of wave propagation. Experiments on an open dataset showed that the average DOA error of the proposed method was 0.528 degrees, and it outperformed a conventional IV-based and DNN-based DOA estimation method.

Motivation & Objective

• Improve the robustness of IV-based DOA estimation in low-SNR and reverberant environments.
• Address the limitations of purely parametric IV-based methods, which degrade under noise and reverberation.
• Integrate DNN-based signal enhancement with physical wave propagation models to preserve interpretability while improving accuracy.
• Demonstrate that combining DNN-based preprocessing with physics-based DOA estimation yields superior performance over end-to-end DNNs or standalone IV methods.
• Validate the effectiveness of spatial augmentation and multi-task learning in enhancing DOA estimation robustness.

Proposed method

• Utilizes first-order ambisonics (FOA) B-format signals to compute time-frequency domain sound intensity vectors (IVs) from four-channel spectrograms.
• Applies a T-F mask (MASKnet) to suppress noise by selecting high-power T-F bins based on energy thresholds.
• Employs a second DNN (RIVnet) to reduce reverberation effects by estimating and subtracting residual interference from IVs.
• Combines the denoised and dereverberated IVs to compute time-series DOA using physics-based formulas: $\phi_t = \arctan(I_{Y,t}/I_{X,t})$, $\theta_t = \arctan(I_{Z,t}/\sqrt{I_{X,t}^2 + I_{Y,t}^2})$.
• Uses a multi-task loss function combining DOA regression loss and sound activity detection (SAD) loss to jointly optimize for angle estimation and source presence.
• Applies FOA-domain spatial augmentation with 16 patterns to improve generalization and robustness during training.

Experimental results

Research questions

• RQ1 Can DNN-based denoising and dereverberation improve the accuracy of IV-based DOA estimation in noisy and reverberant environments?
• RQ2 Does combining DNN-based signal enhancement with physics-based DOA estimation yield better performance than end-to-end DNNs or conventional IV methods?
• RQ3 How effective is the use of a T-F mask (MASKnet) and a residual IV estimation network (RIVnet) in refining IVs for DOA estimation?
• RQ4 To what extent does spatial augmentation in the FOA domain improve the robustness of the proposed method?
• RQ5 Can the integration of SAD with DOA estimation improve overall system performance in time-series DOA tasks?

Key findings

• The proposed method achieved an average DOA error of 0.528°, significantly outperforming the conventional IV-based method and the end-to-end DNN baseline.
• Architecture B (with both MASKnet and RIVnet) achieved lower DOA error than architecture A (without RIVnet), demonstrating the effectiveness of dereverberation in improving IV-based DOA estimation.
• Architecture C, which used FOA-domain spatial augmentation, achieved the lowest average DOA error, confirming that data augmentation enhances DNN generalization in DOA estimation.
• The frame-recall (FR) performance of the proposed method was higher than that of the conventional IV-based method, indicating improved sound activity detection accuracy.
• The multi-task learning setup with joint DOA and SAD loss improved the overall robustness and accuracy of time-series DOA estimation.
• Post-processing via median smoothing of discretized DOA estimates (10° resolution) further improved consistency and reduced angular jitter in event-level DOA outputs.