IPDnet: A Universal Direct-Path IPD Estimation Network for Sound Source Localization

TL;DR

IPDnet targets robust sound-source localization in adverse acoustics by estimating direct-path inter-channel phase differences (DP-IPD) from multichannel STFT signals and mapping them to source locations using known array geometry. It introduces a full-band and narrow-band fusion network, a multi-track DP-IPD learning target for handling multiple sources, and a variable-array design that generalizes to unseen microphone topologies while decoupling DP-IPD estimation from localization. The non-source DP-IPD target is defined via the mean DP-IPD manifold, aided by frame-level activity detection and frame-level PIT training, and localization is performed by matching estimated DP-IPDs to theoretical templates. Experiments on simulated data and LOCATA demonstrate strong localization accuracy and notable generalization across array configurations and elevation estimation, indicating practical potential for robust SSL in real-world environments.

Abstract

Extracting direct-path spatial feature is crucial for sound source localization in adverse acoustic environments. This paper proposes the IPDnet, a neural network that estimates direct-path inter-channel phase difference (DP-IPD) of sound sources from microphone array signals. The estimated DP-IPD can be easily translated to source location based on the known microphone array geometry. First, a full-band and narrow-band fusion network is proposed for DP-IPD estimation, in which alternating narrow-band and full-band layers are responsible for estimating the rough DP-IPD information in one frequency band and capturing the frequency correlations of DP-IPD, respectively. Second, a new multi-track DP-IPD learning target is proposed for the localization of flexible number of sound sources. Third, the IPDnet is extend to handling variable microphone arrays, once trained which is able to process arbitrary microphone arrays with different number of channels and array topology. Experiments of multiple-moving-speaker localization are conducted on both simulated and real-world data, which show that the proposed full-band and narrow-band fusion network and the proposed multi-track DP-IPD learning target together achieves excellent sound source localization performance. Moreover, the proposed variable-array model generalizes well to unseen microphone arrays.

Figures (8)

• Figure 1: Block diagram of the proposed method.
• Figure 2: Examples of non-source target for two different microphone distances.
• Figure 3: Network architecture for the proposed fix-array and variable-array models. The data organization is in the format of: number of sequences $\times$sequence length$\times$ feature dimension.
• Figure 4: Test microphone arrays.
• Figure 5: DP-IPD estimations for the proposed model with or without full-band layers. The acoustic condition for left figure is: RT60 = 0.6 s, SNR = 0 dB (white noise) and for right figure is: RT60 = 0.6 s, SNR = 10 dB (white noise).