On Nonlinear Spatial Filtering in Multichannel Speech Enhancement

Kristina Tesch, Robert Rehr, Timo GerkmannInterspeech 2019 · OpenReview

Abstract

Traditional multichannel speech enhancement relies on linear spatial filters, such as the Minimum Variance Distortionless Response (MVDR) beamformer, which are theoretically optimal under the assumption of Gaussian noise. However, real-world acoustic environments often exhibit "heavy-tailed" (non-Gaussian) noise characteristics. This paper investigates the theoretical performance ceiling of nonlinear spatial filtering by comparing standard linear approaches against an optimal nonlinear Minimum Mean Square Error (MMSE) estimator derived for Gaussian Mixture Model (GMM) noise. Results indicate that nonlinear filtering can provide significant gains—up to 2.6 dB in segmental SNR and 0.5 points in POLQA scores—particularly as the non-Gaussianity of the noise increases.

Introduction

In multichannel speech enhancement, the goal is to utilize spatial information from multiple microphones to isolate a target speaker while suppressing background noise. The mathematical foundation for most current systems is the linear spatial filter. While optimal for Gaussian noise, these filters are sub-optimal for the impulsive, heavy-tailed noise distributions found in real-world settings like cafeterias or busy streets. With the advent of Deep Neural Networks (DNNs) capable of learning complex nonlinearities, it is crucial to quantify the potential benefits of moving beyond linear spatial filtering.

Methodology

The study evaluates three distinct processing frameworks:

  1. MVDR: A baseline linear spatial filter.
  2. MVDR + MMSE Postfilter: A linear spatial filter followed by a single-channel noise reduction stage.
  3. Nonlinear MMSE Estimator: A joint spatial-spectral nonlinear filter optimized for noise modeled by a Gaussian Mixture Model (GMM).

Experiments were conducted using a simulated five-microphone linear array. The "non-Gaussianity" of the noise was controlled via the multivariate kurtosis factor ($q$), testing across three spatial scenarios: spatially white, diffuse, and directional noise. Speech signals were sourced from the WSJ0 dataset, while real-world noise was taken from the CHiME-3 dataset.

Performance improvement in SNR and POLQA scores as a function of the kurtosis factor $q$ for different noise fields.

Key Insights: The Kurtosis Connection

The primary finding is that the advantage of nonlinear filtering is intrinsically linked to the noise distribution's kurtosis. As shown in the experimental results, when noise is nearly Gaussian (low kurtosis), linear filters perform adequately. However, as the noise becomes more heavy-tailed (high kurtosis), the nonlinear MMSE estimator significantly outperforms the linear-plus-postfilter baseline.

A critical distinction is the "inseparability" of optimal nonlinear processing. Unlike linear systems, where spatial filtering (beamforming) and spectral filtering (post-filtering) can be treated as sequential steps, the optimal nonlinear filter must process spatial and spectral information jointly to achieve maximum noise suppression.

Results and Discussion

For stationary noise with high kurtosis, the nonlinear spatial filter achieved an improvement of up to 2.6 dB in segmental SNR over the linear-plus-postfilter setup. Perceptual quality, measured by POLQA scores, also saw a substantial gain of approximately 0.5 points.

Impact of the number of GMM components $M$ on enhancement performance.

The complexity of the noise model, represented by the number of GMM components ($M$), also plays a role in performance. As $M$ increases, the model better captures the non-Gaussian characteristics of the noise, leading to further improvements in both SNR and perceptual quality. On real-world data from the CHiME-3 dataset, the nonlinear approach provided an average improvement of 1.2 dB SNR over linear counterparts when using oracle parameter estimation.

Conclusion

This research demonstrates that real-world noise is sufficiently non-Gaussian to justify the development of nonlinear spatial filters. While the optimal MMSE estimator used here is computationally intensive and relies on oracle information, it establishes a clear performance ceiling. These findings provide a theoretical motivation for designing DNN-based beamformers that can approximate these joint nonlinear functions to achieve superior speech enhancement in complex acoustic environments.

Summary updated

BibTeX
@inproceedings{tesch2019nonlinear,
  author = {Kristina Tesch and Robert Rehr and Timo Gerkmann},
  title = {{On Nonlinear Spatial Filtering in Multichannel Speech Enhancement}},
  booktitle = {Interspeech},
  year = {2019},
  url = {https://openreview.net/forum?id=CeHOv4OPno},
}
Figure