Spatial & Binaural Evaluation

Updated 9 June 2026

Spatial and binaural evaluation is the systematic quantification of interaural cues such as ILD, ITD/IPD, and spatial impression for immersive audio experiences.
The evaluation framework integrates specialized metrics, algorithms, and architectures to preserve spatial fidelity and maintain signal integrity in two-channel audio.
This domain supports practical applications including speech enhancement, music separation, and AR/VR by ensuring accurate, perceptually aligned spatial reproduction.

Spatial and binaural evaluation encompasses the quantification and analysis of perceptually salient cues—such as interaural level difference (ILD), interaural phase/time difference (IPD/ITD), and spatial impression—that govern the sense of source direction and immersion in two-channel (binaural) audio signals. This domain integrates metrics, algorithms, loss functions, and experimental protocols that jointly address signal fidelity, cue preservation, and perceptual realism, with applications spanning speech enhancement, music separation, spatial audio synthesis, and immersive device benchmarking. The following sections delineate foundational principles, standard and emerging evaluation methodologies, representative architectures, benchmark results, and current challenges in spatial and binaural assessment.

1. Spatial Cue Foundations: ILD, ITD/IPD, and Binaural Signal Formulations

Perceptual localization of sound on the horizontal plane is predominantly determined by interaural cues. The interaural level difference (ILD) measures the log-amplitude ratio between left and right channels at frequency $\omega$ : $\mathrm{ILD}(\omega) = 20\log_{10}\frac{|X_L(\omega)|}{|X_R(\omega)|}$ Interaural phase difference (IPD) and its time-domain analog, interaural time difference (ITD), capture time-of-arrival disparities: $\mathrm{ITD}(\omega) = \frac{\Delta\phi(\omega)}{2\pi\,\omega} ,~~ \Delta\phi(\omega) = \angle X_L(\omega) - \angle X_R(\omega)$ Accurate reproduction or preservation of these cues in processed binaural outputs is critical for naturalness, spatial awareness, and externalization.

Binaural enhancement and synthesis models typically operate in the short-time Fourier domain, with various strategies to decouple or jointly estimate target and noise spatial characteristics. For instance, the lightweight LBCCN architecture (see Section 3) predicts frequency-dependent relative acoustic transfer functions (RATF) to reconstruct channels such that phase cues (IPD/ITD) are more faithfully retained than with independent mask-based methods (Wang et al., 2024).

2. Evaluation Metrics for Spatial and Binaural Fidelity

Spatial evaluation metrics fall into two main classes: (a) direct measurement of binaural cue preservation and (b) perceptual or task-oriented similarity scoring.

Binaural Cue Preservation Metrics:

ILD-error (dB):

$\mathrm{ILD\text{-}error} = \Bigl|\,20\log_{10}\frac{|X_L|}{|X_R|} - 20\log_{10}\frac{|\hat{X}_L|}{|\hat{X}_R|}\Bigr|$

IPD-error (radians):

$\mathrm{IPD\text{-}error} = \bigl|\angle X_L - \angle X_R - (\angle\hat{X}_L - \angle\hat{X}_R)\bigr|$

Spatial Fidelity and Perceptual Metrics:

Modified Binaural STOI (MBSTOI): Intelligibility metric ranging [0,1].
ΔPESQ: Difference in perceptual evaluation of speech quality between processed and noisy signals.
SPL Distance (“Spatial Perception” metric): Time-aggregated Euclidean difference between left-right sound-pressure levels; quantifies both lateralization magnitude and direction (Li et al., 2023).
Signal-to-Spatial Distortion Ratio (SSR) and Signal-to-Residual Distortion Ratio (SRR): Used to isolate spatial from non-spatial degradation (Namballa et al., 30 Jun 2025).
Feature-based Metrics: NSIM-based BINAQUAL (Localization Similarity) (Panah et al., 17 May 2025) and DOA-driven DPLM (Manocha et al., 2021) directly model perceptual localization similarity between reference and processed binaural pairs.

Specialized Maps and Diagnostics:

Error Maps (3DAE): Frequency-time error visualization for magnitude, ILD, IPD, temporal alignment, and loudness (Xu et al., 28 May 2026).

3. Architectural Methodologies for Spatial Cue Preservation

A spectrum of neural and algorithmic architectures has emerged for spatial and binaural audio modeling, each with distinct evaluation implications:

RATF-based Networks: LBCCN leverages explicit RATF prediction, yielding strong phase-cue preservation with minimal parameter and compute cost (38K params, 0.216G MACs, RTF=0.054) (Wang et al., 2024).
Complex Masking Networks: Approaches such as BCCTN employ complex ratio masks for each channel, penalized by ILD and IPD errors directly in loss design (Tokala et al., 2024).
Hybrid Codec Architectures: MAD encodes content versus spatial (IR) cues separately, reconstructing accurate ITD/ILD at high compression rates (Ratnarajah et al., 2023).
Spatial Perception-Driven Generators: Models such as SAGM define new SPL-difference metrics to align generation with time-varying spatial perception (Li et al., 2023).
Parametric/Auditory-Model-Guided Correction: “SpatialNet” integrates an auditory-model-based loss, including interaural vector strength (IVS), to robustly mitigate spatial artifacts under dynamic head rotations (Shamay et al., 23 Dec 2025).
Signal Matching and Decomposition: BSM and its time-frequency decomposed variants optimize filter sets for left/right signals, with error quantified through NMSE, ILD/ITD error, and spectral distortion (Berger et al., 2023).

4. Benchmark Protocols and Comparative Results

Evaluation protocols are tailored to the application scenario:

Speech Enhancement and Hearing Devices: Fixed-source Librispeech signals spatialized via HRTF, with diffuse noise fields and random SNR sampling, form the basis for large-scale, controlled SCP benchmarks (e.g., LBCCN: 50,000 2-s samples, 8:1:1 split, SNRs [–10,10] dB) (Wang et al., 2024).
Binaural Music Separation: Binauralized MUSDB18-HQ (by HRIR convolution) enables quantification of SSR, SRR, and explicit ITD/ILD error per instrument (Namballa et al., 30 Jun 2025).
Real-World and Synthetic Soundfields: Simulated and measured room setups (with diverse arrays, orientations, and BRIRs) anchor objective (e.g., MUSHRA, MAE ERB-averaged ILD/ITD, IACC) and subjective (forced-choice localization, MOS) evaluations (Pawlak et al., 2024).

Representative tabulated results (averages over input SNR –10 to +10 dB for four methods):

Metric	DBSEnh	BiTasNet	BCCTN	LBCCN
MBSTOI ↑	0.85	0.89	0.90	0.93
ΔPESQ ↑	0.06	0.61	0.86	0.78
ILD-error ↓	4.22 dB	4.09 dB	2.27 dB	2.53 dB
IPD-error ↓	0.64 rad	0.83 rad	0.58 rad	0.50 rad
Params	10.5M	1.7M	11.1M	38K
RTF	0.022	0.329	0.228	0.054

LBCCN achieves strong NR (MBSTOI, ΔPESQ) and SCP (ILD/IPD-error) commensurate with or surpassing larger DNNs, at a fraction of complexity (Wang et al., 2024).

5. Subjective and Objective Hybrid Assessments

While objective metrics quantify local or global binaural fidelity, listening tests are critical for perceptual validation.

MUSHRA (BS.1534-1): Used for perceptual spatial and timbral fidelity in room auralization and BSM correction evaluation; SDM+center/HO-SIRR achieve medians ≈4/5, anchor ≈1, reference 5 (Pawlak et al., 2024); post-processed SpatialNet models matched reference for θ_rot=60–90° (Shamay et al., 23 Dec 2025).
Stereo Preference and Forced Localization: PseudoBinaural achieves comparable subjective spatial impression to fully supervised and ground-truth baselines, corroborated by a newly introduced phase-difference metric highly sensitive to L/R cue errors (Xu et al., 2021).
Time-Varying/Visual-Cued Tests: SPL-Distance metric supports longitudinal spatial evaluation without requiring user panels, but visually-cued models (e.g., ViSAudio) also employ expert MOS on spatial impression, consistency, alignment, and realism (Li et al., 2023, Zhang et al., 2 Dec 2025).

6. Challenges, Best Practices, and Open Problems

Persistent challenges in spatial and binaural evaluation include:

Metric Sensitivity and Coverage: Classical metrics (ILD/ITD error) may fail under reverberant, non-stationary, or multi-source scenes. Hybrid feature-driven (BINAQUAL, DPLM) and error-map visualizations (3DAE) address this by highlighting localized or perceptually salient error modes (Panah et al., 17 May 2025, Manocha et al., 2021, Xu et al., 28 May 2026).
Interplay of NR and SCP: There is an inherent trade-off between aggressive noise reduction and spatial-cue fidelity; direct RATF regression and explicit binaural loss terms mitigate this but require careful loss scheduling and network design (Wang et al., 2024, Tokala et al., 2024, Shamay et al., 23 Dec 2025).
Resource Constraints: Real-time evaluation on low-power hardware (e.g., hearing aids) demands highly efficient models and metrics compatible with limited compute budgets (Wang et al., 2024).
Subjectivity and Perceptual Alignment: No metric wholly predicts listener preference or externalization; best practice combines full-reference objective scores (MBSTOI, ILD/IPD-error, NSIM/DPLM) with focused listening studies.
Generalization: Robust evaluation must stress models with variable HRTFs, source types, and listener/head motion, using both controlled and cross-dataset tests (Xu et al., 2021, Shamay et al., 23 Dec 2025).

Future work emphasizes metric standardization, rigorous benchmark design (including open-source pipelines), and refined perceptual modeling—especially integrating listener-specific HRTFs, elevation sensitivity, and dynamic scene adaptation (Panah et al., 17 May 2025, Shamay et al., 23 Dec 2025, Xu et al., 28 May 2026).

7. Impact and Emerging Directions

Spatial and binaural evaluation is now fundamental to development and deployment in assistive hearing, AR/VR, binaural codecs, and audio synthesis. The field is witnessing a transition from separate, cue-based signal processing methods to end-to-end architectures guided by domain-informed, perceptually aligned, full-reference metrics. Recent advances extend spatial benchmarks to language-guided synthesis, visually-cued inference, and multi-turn localization reasoning (Biswas et al., 30 Sep 2025, Pan et al., 1 Jun 2025). Diagnostic error mapping and hybrid objective–subjective frameworks are enhancing transparency and interpretability for both research and practical deployment (Shamay et al., 23 Dec 2025, Xu et al., 28 May 2026). As standards evolve and spatial audio applications proliferate, rigorous, interpretable spatial and binaural evaluation will remain a technical cornerstone (Wang et al., 2024, Panah et al., 17 May 2025, Xu et al., 28 May 2026).