Skip to content
Open access

Spatial Self-Attention Driven Feature Re-Weighting for Sound Source Localization in Challenging Acoustic Environments

2026 · IEEE Access · Vol 14, pp. 164871-164886 · 0 citations · 46 references

Abstract

Sound source localization remains vulnerable in challenging acoustic environments characterized by reverberation and low signal-to-noise ratios. To address this issue, this paper presents spatial self-attention-driven feature re-weighting within a convolutional neural network (SSA-CNN). The re-weighting mechanism combines channel-wise average pooling and channel-wise max pooling to derive complementary spatial statistics and then generates an adaptive weight mask for acoustic feature maps. A training dataset is constructed from the LibriSpeech corpus, and comparative experiments are conducted against GCC-PHAT, SRP-PHAT, TFnet, SELDnet, and MSAT under controlled acoustic conditions and LOCATA-based evaluation. Experimental results show that SSA-CNN maintains a root mean squared angular error (RMSAE) of 7°–15° across multiple reverberation-time (T60) and signal-to-noise-ratio (SNR) settings in the controlled tests. Under the challenging condition T $60=1.5$ s and SNR = 5 dB, SSA-CNN achieves an error of 23.09°, which is lower than the errors obtained by the evaluated baselines under the adopted protocol. These results demonstrate the effectiveness of spatial self-attention-driven feature re-weighting in improving the accuracy and interference robustness of three-dimensional sound source localization.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.