# Optimal Transport Audio Distance with Learned Riemannian Ground Metrics

Source: [Optimal Transport Audio Distance with Learned Riemannian Ground Metrics](<https://www.sota2.com/research/paper/69fd6df65497b3b670c8933a>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-09T05:00:04.795Z

Authors: Wonwoo Jeong

Published: 2026-05-07T01:01:18.000Z

## Abstract

In audio generation evaluation, Fr\\'echet Audio Distance (FAD) is a 2-Wasserstein distance with structural constraints for both primitives: the cost is a frozen embedding pullback whose invariance set hides severe artifacts, and the coupling is a Gaussian fit that dilutes rank-1 contamination relative to discrete OT. We propose Optimal Transport Audio Distance (OTAD), which corrects each primitive with one dedicated mechanism -- a residual Riemannian ground-metric adapter for the cost and entropic Sinkhorn optimal transport for the coupling. Across eight encoders under a four-axis protocol, coupling-only comparisons at $\\epsilon = 0.05$ show that Sinkhorn's rank-1 sensitivity exceeds FAD's by a factor of 1.9 to 3.6. Furthermore, OTAD achieves a higher mean Spearman correlation with audio-quality MOS (DCASE 2023 Task 7) than baseline metrics. As an intrinsic benefit of the discrete transport plan, OTAD yields per-sample diagnostics with AUROC $\\ge 0.86$, a capability that scalar- or kernel-aggregated metrics structurally lack.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Audio quality assessment on DCASE Task 7 System-level n=9 2023](<https://www.sota2.com/research/sota/audio-quality-assessment-on-dcase-task-7-system-level-n-9-2023.md>): Audio Quality Assessment; DCASE Task 7 System-level n=9 2023; FAD: not recorded.

- [Audio Quality Assessment on DCASE Task 7 Per-category granularity 2023](<https://www.sota2.com/research/sota/audio-quality-assessment-on-dcase-task-7-per-category-granularity-2023.md>): Audio Quality Assessment; DCASE Task 7 Per-category granularity 2023; FAD: not recorded.

- [Artifact Detection on Per-sample diagnostics Gaussian noise](<https://www.sota2.com/research/sota/artifact-detection-on-per-sample-diagnostics-gaussian-noise.md>): Artifact Detection; Per-sample diagnostics Gaussian noise; AUROC: not recorded.

- [Artifact Detection on Per-sample diagnostics Cross-class noise](<https://www.sota2.com/research/sota/artifact-detection-on-per-sample-diagnostics-cross-class-noise.md>): Artifact Detection; Per-sample diagnostics Cross-class noise; AUROC: not recorded.

- [Artifact Detection on Per-sample diagnostics Silence noise](<https://www.sota2.com/research/sota/artifact-detection-on-per-sample-diagnostics-silence-noise.md>): Artifact Detection; Per-sample diagnostics Silence noise; AUROC: not recorded.

- [Audio category-fit MOS correlation assessment on DCASE Task 7 2023 (test)](<https://www.sota2.com/research/sota/audio-category-fit-mos-correlation-assessment-on-dcase-task-7-2023-test.md>): Audio category-fit MOS correlation assessment; DCASE Task 7 2023 (test); FAD: not recorded.

- [Audio Quality Assessment on DCASE Task 7 Per-category n=63 2023](<https://www.sota2.com/research/sota/audio-quality-assessment-on-dcase-task-7-per-category-n-63-2023.md>): Audio Quality Assessment; DCASE Task 7 Per-category n=63 2023; FAD: not recorded.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.05554>)

[PDF](<https://arxiv.org/pdf/2605.05554.pdf>)