Acoustics: Spatial Properties

Emmanuel Vincent, Sharon Gannot, Tuomas Virtanen2018 · DOI

Abstract

In multichannel scenarios, the spatial distribution of sound sources provides critical cues for signal separation and enhancement. This paper explores the foundational elements of acoustics and sound engineering used to model multichannel mixtures. We examine the relationship between source signals and their spatial images across different channels, considering both physical microphone recordings and artificial mixtures. We introduce several modeling approaches, including Acoustic Transfer Functions (ATF), Relative Transfer Functions (RTF), and full-rank spatial covariance models, linking these parameters to the physical geometry of the sound scene and the source direction of arrival.

Introduction

While single-channel audio processing relies primarily on spectral cues such as pitch and timbre, multichannel systems exploit the spatial characteristics of sound. Modeling these characteristics is challenging due to the complexity of real-world environments where sound undergoes reflections (reverberation), diffraction around obstacles, and absorption by various surfaces. This paper formalizes these physical phenomena into mathematical models suitable for algorithmic source separation and speech enhancement, transitioning from idealized free-field environments to complex indoor spaces.

Modeling Framework

The General Mixing Model

Spatial modeling begins with a two-stage process: individual source signals are first "spatialized"—transformed into multichannel images—and then combined. In professional audio production, this may involve nonlinear post-mixing operations, but for many signal processing tasks, linear approximations are preferred.

Linear Time-Invariant (LTI) Modeling

When sources and microphones are static, the mixing process is modeled as a convolution with Acoustic Impulse Responses (AIRs). An AIR typically consists of three distinct temporal components:

  1. Direct Path: The initial peak representing the line-of-sight travel.
  2. Early Echoes: Distinct, sparse reflections from nearby surfaces.
  3. Late Reverberation: A dense, exponentially decaying tail of reflections.
Acoustic Impulse Responses (AIRs) for two microphones, showing the direct path, early reflections, and late reverberation tail.

The distinction between these components is vital. For instance, the "mixing time" boundary ($\tau_c$), often around 50ms, marks the transition from sparse early reflections to a statistical "diffuse" late reverberation field, which is a key parameter for dereverberation algorithms.

Frequency Domain and Narrowband Approximations

To facilitate efficient computation, models often move to the Short-Time Fourier Transform (STFT) domain. This transforms complex time-domain convolutions into simpler complex-valued multiplications. Key tools in this domain include:

  • Steering Vectors: These encode the direction of arrival (DOA) and distance.
  • Relative Transfer Functions (RTFs): These model the relationship between channels, focusing on interchannel differences rather than absolute paths.
  • Interchannel Cues: Formalized metrics for Interchannel Level Differences (ILD), Phase Differences (IPD), and Time Differences (ITD).

Key Insights in Spatial Acoustics

Interchannel Coherence (IC) and Spatial Aliasing

A significant finding in indoor acoustics is that the Interchannel Coherence (IC) of the reverberant field depends primarily on microphone spacing and frequency, rather than specific room characteristics. However, spatial modeling faces limits such as "spatial aliasing." Interchannel Time Difference (ITD) becomes ambiguous above a certain frequency determined by microphone spacing, necessitating the integration of phase information across multiple frequencies to accurately recover source positions.

From Vectors to Matrices

As environments become more reverberant, simple rank-1 steering vectors often fail to capture the "smeared" spatial image. The shift to second-order statistical models, specifically full-rank spatial covariance matrices, allows for better handling of long impulse responses and diffuse noise sources. In many cases, the spatial covariance matrix converges to a parametric model combining a rank-1 direct-path component with a diffuse-field coherence matrix.

Conclusion

Understanding the spatial properties of acoustics is essential for modern multichannel signal processing. By bridging physical acoustics with mathematical modeling—from simple LTI filters to complex covariance matrices—researchers can develop more robust algorithms for unmixing overlapping sound sources. Future work points toward data-driven "manifold learning" to capture the complex variations of impulse responses that purely geometric models may miss.

Summary updated

Abstract

In multichannel scenarios, the fact the acoustic scene is observed from different positions in space can be exploited for separation and enhancement. In this chapter, we recall basic elements of acoustics and sound engineering, and use them to model multichannel mixtures. We consider the general relationship between a source signal and its spatial image in a given channel, and examine how it translates in the case of microphone recordings or artificial mixtures. We then introduce several possible models based on the acoustic transfer function, the relative transfer function and interchannel phase and level differences, and a full-rank spatial covariance model. We link the parameters of these models to the geometry of the sound scene, including the source direction of arrival.
BibTeX
@misc{vincent2018acoustics,
  author = {Emmanuel Vincent and Sharon Gannot and Tuomas Virtanen},
  title = {{Acoustics: Spatial Properties}},
  year = {2018},
  doi = {10.1002/9781119279860.ch3},
}
Figure