Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust DOA estimation using deep acoustic imaging

Adrian S. Roman, Iran R. Roman, Juan P. Bello

TL;DR

The paper addresses DoAE using SIMs (spherical intensity maps) as input and tackles the limitation of low-resolution microphone arrays by introducing a complex-valued DBPN-based upsampling from 4ch to 32ch. It benchmarks SIM-based DoAE with DeepWave, incorporating a K-means post-processing pipeline and GRU-enhanced end-to-end models with ADPIT loss, validating on LOCATA and STARSS23 datasets. The results show that 32ch SIM-based DoAE can achieve competitive localization performance (e.g., $LE \,=\$ $14.8^\8$ and $LR \,=\$ $99.20$ on LOCATA), while GRU variants improve LE further, albeit with trade-offs in LR; upsampling 4ch inputs maintains potential when integrated into a DeepWave-like pipeline. Overall, the work demonstrates the relevance and practicality of acoustic imaging for DoAE and provides a scalable path to leverage SIMs on datasets with varying microphone counts.

Abstract

Direction of arrival estimation (DoAE) aims at tracking a sound in azimuth and elevation. Recent advancements include data-driven models with inputs derived from ambisonics intensity vectors or correlations between channels in a microphone array. A spherical intensity map (SIM), or acoustic image, is an alternative input representation that remains underexplored. SIMs benefit from high-resolution microphone arrays, yet most DoAE datasets use low-resolution ones. Therefore, we first propose a super-resolution method to upsample low-resolution microphones. Next, we benchmark DoAE models that use SIMs as input. We arrive to a model that uses SIMs for DoAE estimation and outperforms a baseline and a state-of-the-art model. Our study highlights the relevance of acoustic imaging for DoAE tasks.

Robust DOA estimation using deep acoustic imaging

TL;DR

The paper addresses DoAE using SIMs (spherical intensity maps) as input and tackles the limitation of low-resolution microphone arrays by introducing a complex-valued DBPN-based upsampling from 4ch to 32ch. It benchmarks SIM-based DoAE with DeepWave, incorporating a K-means post-processing pipeline and GRU-enhanced end-to-end models with ADPIT loss, validating on LOCATA and STARSS23 datasets. The results show that 32ch SIM-based DoAE can achieve competitive localization performance (e.g., and on LOCATA), while GRU variants improve LE further, albeit with trade-offs in LR; upsampling 4ch inputs maintains potential when integrated into a DeepWave-like pipeline. Overall, the work demonstrates the relevance and practicality of acoustic imaging for DoAE and provides a scalable path to leverage SIMs on datasets with varying microphone counts.

Abstract

Direction of arrival estimation (DoAE) aims at tracking a sound in azimuth and elevation. Recent advancements include data-driven models with inputs derived from ambisonics intensity vectors or correlations between channels in a microphone array. A spherical intensity map (SIM), or acoustic image, is an alternative input representation that remains underexplored. SIMs benefit from high-resolution microphone arrays, yet most DoAE datasets use low-resolution ones. Therefore, we first propose a super-resolution method to upsample low-resolution microphones. Next, we benchmark DoAE models that use SIMs as input. We arrive to a model that uses SIMs for DoAE estimation and outperforms a baseline and a state-of-the-art model. Our study highlights the relevance of acoustic imaging for DoAE tasks.
Paper Structure (13 sections, 1 equation, 3 figures, 2 tables)

This paper contains 13 sections, 1 equation, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Approach
  4. DoAE from acoustic images via K-means clustering
  5. Upsampling $\mathbf{\Sigma}$ using super-resolution
  6. DeepWave SIMs as inputs for DoAE
  7. Methodology
  8. Datasets
  9. Models, ablations, and baselines
  10. Training procedure and evaluation metrics
  11. Results
  12. Conclusion and future work
  13. Acknowledgements

Figures (3)

  • Figure 1: The acoustic image by DeepWave (top) and DASB (bottom) for two sound sources. Dots denote ground truth.
  • Figure 2: Signal pipeline in the models we study.
  • Figure 3: CDBPN upsampling of a single reverberant white noise source directly facing the front of the microphone.