Spherical Harmonic Representations for Spatial Audio

41 concepts30 papers1 proposal

Imagine trying to capture and reproduce a complex 3D soundscape. The traditional approach relies on raw multi-channel audio, recording signals from a specific configuration of microphones. However, this method tightly couples the audio data to the physical hardware layout. If you change the microphone array geometry or attempt playback on a different loudspeaker setup, the spatial relationships collapse.

To overcome this limitation, modern spatial audio engineering shifts its perspective: instead of treating a soundfield as a collection of discrete, point-wise channel signals, it treats the soundfield as a continuous mathematical function distributed across the surface of a sphere. By applying the Spherical Harmonic Transform (SHT), we can decompose a 3D acoustic environment into a set of orthogonal basis functions called spherical harmonics. This is conceptually identical to how a Fourier transform decomposes a time-domain signal into discrete sine waves, but executed over a angular, spherical domain.

Decoupling Acoustic Geometry from Hardware

The core intuition behind processing in the spherical harmonic domain—frequently manifested as Ambisonics—is that it provides a coordinate-independent, device-agnostic representation of the soundfield.

As detailed by Sarabia et al., transforming multi-channel recordings into spherical harmonic coefficients allows machine learning models to learn intrinsic spatial features. Because these features describe the soundfield itself rather than the voltages on specific microphone capsules, models generalize far better across entirely different hardware setups.

Raw Multi-Channel Input (Device-Dependent)
       │
       ▼ [Spherical Harmonic Transform]
Spherical Harmonic Coefficients (Device-Agnostic Soundfield)
       │
       ├───────────────────────────┤
       ▼                           ▼
B-Format / Ambisonics       Higher-Order Spatial Upscaling
(Robust Localization)       (Sharper Spatial Resolution)

This representation scales gracefully. A soundfield can be compressed into First-Order Ambisonics (FOA) for efficient processing, which Guo et al. demonstrate provides a highly robust mathematical framework for sound event localization and detection. When greater fidelity is required, the representation scales up to Higher-Order Ambisonics (HOA). As explained by Narbutt et al., adding higher-order coefficients acts like adding high-frequency terms in a Fourier series, sharpening the spatial resolution and allowing for highly precise localization of sound sources. Because the underlying representation is continuous, this precise soundfield can be decoded perfectly onto any arbitrary playback speaker configuration.

Separating Direction from Frequency

Why does transforming audio into spherical harmonics simplify spatial signal processing? In raw multi-channel audio streams, spatial information (the angle of arrival) and spectral information (the frequency) are deeply entangled across channels.

Decomposing the soundfield into the spherical harmonic domain cleanly separates these variables. As shown by Tao et al., this decomposition isolates angular distribution from frequency information, allowing algorithms to perform spatial filtering and noise suppression on specific directional modes without distorting the underlying audio frequency spectrum.

By mapping the acoustic environment directly to the geometry of a sphere, algorithms can resolve spatial ambiguities that plague raw time-domain signals. For example, Singhal et al. leverage this domain to construct partial dictionaries that pinpoint the direction of arrival for multiple overlapping speech sources. Similarly, Evers et al. demonstrate that processing acoustic signals in the spherical harmonic domain captures the physical geometry of the sound field more effectively than standard time-domain methods, significantly improving the tracking of moving speakers in highly reverberant rooms.

Compact Encodings for Deep Learning and Simulation

Simulating how sound waves scatter, diffract, and bounce off complex 3D objects is computationally expensive. If a neural network or graphics engine attempts to predict a dense, continuous 3D grid of acoustic pressure values over time, it quickly succumbs to the curse of dimensionality.

Spherical harmonics serve as an exceptionally powerful compression mechanism for spatial data. Because acoustic scattering patterns vary smoothly over angles, the entire continuous angular distribution of a scattered wave can be compactly encoded into a handful of low-order spherical harmonic coefficients.

Tang et al. and Meng et al. utilize this exact property to achieve real-time sound propagation in dynamic virtual environments. Instead of forcing a neural network to output high-dimensional spatial data, they train models to output a small set of spherical harmonic coefficients that describe the scattering field. This highly compressed representation allows complex, wave-based phenomena like diffraction to be injected directly into real-time geometric ray-tracing engines with virtually zero computational overhead.

This principle applies equally to speech enhancement. By transforming multi-channel microphone signals into the spherical harmonic domain, Pan et al. inject spatial geometry directly into a dual-encoder network, allowing the model to extract clean speech without the heavy computational burden of processing raw, uncompressed multi-channel audio streams.

Global Control via Modal Signal Processing

When spatial audio is framed as a continuous function on a sphere, complex spatial operations transform into elegant algebraic manipulations.

Consider "directional sharpening"—the process of creating an acoustic spotlight to amplify sound coming from a specific direction while suppressing background noise. Kleijn demonstrates that by treating spatial audio as a field on a spherical surface, the multiplication of two distinct spherical harmonic expansions can be computed efficiently as a matrix multiplication using Clebsch-Gordan coefficients. This allows low-degree Ambisonic signals to be upscaled into high-degree, highly directional representations with minimal timbre distortion.

Point-Wise Control (Traditional)       Mode-Domain Control (Spherical Harmonics)
     [·]      [·]      [·]                       ┌───────────────┐
                                                 │  Y_0^0  Y_1^-1│
     [·]      [M]      [·]                       │               │
                                                 │  Y_1^0  Y_1^1 │
     [·]      [·]      [·]                       └───────────────┘
Cancels noise only at discrete             Cancels noise continuously across
microphone coordinates [·].                the entire harmonic region.

This "mode-domain" perspective fundamentally changes fields like Active Noise Control (ANC). Traditional ANC systems attempt to cancel noise point-by-point at discrete microphone locations. In contrast, Maeno et al. model the entire acoustic environment as a summation of harmonics rather than discrete spatial points. This allows for the continuous control of sound properties across an entire defined region. By operating entirely within the spherical harmonic domain, Maeno et al. combine sparse decomposition with the acoustic addition theorem to cancel incoming noise globally across an entire spatial zone using distributed, non-concentric sensor arrays.

Key Foundational Work

A sparse reconstruction method for speech source localization using partial dictionaries over a spherical microphone array (2014) — contributed 1 concept elements to this direction.
Capture and Recreation of Higher-Order 3D Sound Fields via Reciprocity (2004) — contributed 1 concept elements to this direction.
Semi-Analytical Computation of Acoustic Scattering by Spheroids and Disks (2014) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_spherical_harmonic_domain_processing_c08acfe8,
  title = {spherical harmonic domain processing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spherical-harmonic-domain-processing/rdr_c08acfe8379c4e49a754277fdd0bffeb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_spherical_harmonic_domain_processing_c08acfe8_v1,
  title = {spherical harmonic domain processing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/spherical-harmonic-domain-processing/rdr_c08acfe8379c4e49a754277fdd0bffeb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake