Pseudo-Determined Blind Source Separation for Ad-hoc Microphone Networks

Abstract

We propose a pseudo-determined blind source separation framework that exploits the information from a large number of microphones in an ad-hoc network to extract and enhance sound sources in a reverberant scenario. After compensating for the time offsets and sampling rate mismatch between (asynchronous) signals, we interpret as a determined M × M mixture the over-determined M × N mixture, where M > N is the number of microphones and N is the number of sources. Next, we propose a pseudo-determined mixture model that can apply an M ×M independent component analysis (ICA) directly to the M-channel recordings. Moreover, we propose a reference-based permutation alignment scheme that aligns the permutation of the ICA outputs and classifies them into target channels, which contain the N sources, and non-target channels, which contain reverberation residuals. Finally, using the signals from non-target channels, we estimate in each target channel the power spectral density of the noise component that we suppress with a spectral post-filter. Interestingly, we also obtain latereverberation suppression as by-product. Experiments show that each processing block improves incrementally source separation and that the performance of the proposed pseudo-determined separation improves as the number of microphones increases.

A Blind Source Separation Framework for Ego-Noise Reduction on Multi-Rotor Drones

“…While ICA-based BSS can suppress directional ego-noise effectively, there are still several issues that remain unsolved when using BSS in practice. First, permutation ambiguity becomes a crucial and challenging problem in low-SNR scenarios, especially when the microphones outnumber the sources, leading to an over-determined mixture [56], [57]. Second, BSS typically works as a batch process and thus requires the acoustic mixing network to remain stationary for a certain interval, i.e.…”

“…We apply an M × M ICA directly to the M -channel input, assuming an M × M mixing network with M sources [57]. These M sources contain a target sound source componentS and…”

“…The proposed method assumes that the microphones M outnumber the sources N , which include the target sound sources and the ego-noise sources. This over-determined case can be treated as a pseudo-determined case [57], where an M × M ICA is applied to separate the sources in the mixture. In practice, an under-determined case (N > M ) may occur and the proposed method might not be able to deal with this scenario robustly, as the condition of formulation an M × M ICA does no longer hold.…”

Distributed Sampling Rate Offset Estimation Over Acoustic Sensor Networks Based on Asynchronous Network Newton Optimization

“…They can constitute an ad-hoc acoustic sensor network, which vastly increases the amount of spatial information compared with the traditional single microphone (or microphone array). Therefore, it boosts the performance of many audio processing tasks, e.g., speech enhancement [1], source separation [2], [3], speaker localization and tracking [4]- [6]. In such a network, each node possesses its independent clock system, resulting in asynchronous sampling rates among nodes.…”

Data Augmentation for Speech Separation

“…In comparison to multi-channel approaches that exploit the spatial information of sound sources (Wang, 2014;Wang and Cavallaro, 2018), single-channel speech separation is a more challenging task (Wang and Chen, 2018). Deep neural networks (DNN) are at the forefront for speech separation and can be broadly categorized into time-frequency domain approaches (Kolbaek et al, 2017;Hershey et al, 2016;Wang et al, 2018b;Williamson et al, 2015;Wang et al, 2022) and timedomain approaches (Luo and Mesgarani, 2018;Luo and Nima Mesgarani, 2019;Chen et al, 2020;Nachmani et al, 2020).…”