ACES: Accent Subspaces for Coupling, Explanations, and Stress-Testing in Automatic Speech Recognition

Swapnil Parekh

cs.SDFeb 28, 2026

ACES: Accent Subspaces for Coupling, Explanations, and Stress-Testing in Automatic Speech Recognition

Authors: Swapnil Parekh

Organizations: Intuit

Abstract

ASR systems exhibit persistent performance disparities across accents, but whether these gaps reflect superficial biases or deep structural vulnerabilities remains unclear. We introduce ACES, a three-stage audit that extracts accent-discriminative subspaces from ASR representations, constrains adversarial attacks to them, and tests whether removing them improves fairness. On Wav2Vec2-base with seven accents, imperceptible perturbations (~60 dB SNR) along the accent subspace amplify the WER disparity gap by nearly 50% (21.3->31.8 pp), exceeding random-subspace controls; a permuted-label test confirms specificity to genuine accent structure. Partially removing the subspace worsens both WER and disparity, revealing that accent-discriminative and recognition-critical features are deeply entangled. ACES thus positions accent subspaces as powerful fairness-auditing tools, not simple erasure levers.

Figures & tables

Figure 1: ACES: subspace extraction, stress-test (waveform PGD, L2 $\varepsilon$ ), and project-out. Conceptually, the representation $\mathbf{h}$ is projected onto the plane spanned by $\mathbf{U}$ ; coupling $m(x)$ measures how much the attack moves $\mathbf{h}$ along that subspace.

Item	Value
Model	Wav2Vec2-base-960h [ 11 ]
Data	Common Voice (en.) [ 23 ] , 7 acc., 100 utt./acc., 700 total
$\varepsilon$ , steps	0.01, 80 (waveform L2)
Layer $\ell^{*}$ , $k$	3, 8 (ridge probe)
Probe acc. (projected)	96.3%
Stability (principal angle)	$\sim$ 50°

Table 1: Setup & subspace: model, data,

\varepsilon

/steps, chosen

\ell^{*}

\mathbf{U}

, probe acc., stability.

Figure 2: Three-track diagnostic (1-column): probe accuracy, corr(projection, WER), stability (principal angle, °) vs. layer. Layer 3 maximizes probe accuracy while maintaining stability below 50°; $k{=}8$ (dashed).

Figure 3: Epsilon ablation: disparity (left) and mean WER (right) vs. perturbation budget

\varepsilon

. The accent-subspace (blue) maintains high disparity as

\varepsilon

grows, while the random control (orange) drops off—the gap widens from +1.8 pp (

\varepsilon{=}0.005

) to +4.8 pp (

\varepsilon{=}0.05

Figure 4: Coupling $m(x)$ vs. $\Delta\text{WER}$ at layer $\ell^{*}$ . Accent-subspace attacks (blue) produce ${\sim}2\times$ higher mean $m(x)$ than random-subspace (orange), confirming preferential activation of accent directions. Per-utterance $r$ is weak for both due to CTC decoding noise.

Accent	Clean	Unconstr.	Random	Accent-sub.
African	27.1	38.0	34.3	34.2
Bermuda	30.9	46.4	42.6	44.9
Indian	40.7	56.7	54.1	56.3
Malaysia	39.8	58.1	51.1	52.6
Singapore	31.9	43.6	40.8	42.2
US	19.4	28.4	23.2	24.5

Table 2: Mean WER (%) by accent and condition;

\varepsilon{=}0.01

, 80 steps,

\beta{=}2

. Disparity = max

-

min. Lowest WER per column in bold .

	Clean		Attacked
Accent	Base	Int.	Base	Int.
African	27.1	27.5	31.6	31.5
Bermuda	30.9	30.8	41.2	42.0
Indian	40.7	41.0	49.0	49.0
Malaysia	39.8	40.4	48.4	48.4
Singapore	31.9	32.2	38.4	38.9

Table 3: Project-out intervention (

\alpha{=}0.5

): WER (%) before (Base) and after (Int.) attenuating the accent subspace on clean and accent-subspace-attacked audio. Attacked WER here uses per-utterance PGD; batch-level aggregation in Table 2 yields slightly different values.

Demographic Bias in ASR

Demographic bias in automatic speech recognition (ASR) concerns unequal transcription accuracy and user burdens across speaker groups and language varieties.

See all 30 papers

Contrastive Regularization for Accent-Robust ASR
May 5, 2026Van-Phat Thai, Aradhya Dhruv, Duc-Thinh Pham +1Automatic Speech Recognition Supervised Contrastive Learning
Air Traffic Management Research Institute, Nanyang Technological University, Singapore · Center of AI Research, VinUniversity, Vietnam
When Entanglement Lower-Bounds Disparity: Auditing and Repairing Demographic Fairness in Audio Understanding Models
Sep 23, 2026Kian Shamsaie, Iman ModarressiDisentangled Representation Learning Audio Representation Learning
People Make Things
Beyond Word Error Rate: Auditing the Diversity Tax in Speech Recognition through Dataset Cartography
Mar 5, 2026Ting-Hui Cheng, Line H. Clemmensen, Sneha DasASR Evaluation Demographic Bias in ASR
Department of Applied Mathematics and Computer Science, Technical University of Denmark, Denmark · Department of Mathematical Sciences, University of Copenhagen, Denmark

May 5, 2026J/K move · Enter open · S save

Contrastive Regularization for Accent-Robust ASR

Van-Phat Thai, Aradhya Dhruv, Duc-Thinh Pham, Sameer Alam

Air Traffic Management Research Institute, Nanyang Technological University, Singapore · Center of AI Research, VinUniversity, Vietnam

Automatic Speech Recognition Supervised Contrastive Learning Robust ASR

Open paper AskPDF arXiv

ASR systems based on self-supervised acoustic pretraining and CTC fine-tuning achieve strong performance on native speech but remain sensitive to accent variability. We investigate supervised contrastive learning (SupCon) as a lightweight, accent-invariant auxiliary objective for CTC fine-tuning. An utterance-level contrastive loss regularizes encoder representations without architectural modification or explicit accent supervision. Experiments on the L2-ARCTIC benchmark show consistent WER reductions across multiple pretrained encoders, with up to 25 -- 29% relative reduction under unseen-accent evaluation. Analysis using within-transcript cosine dispersion indicates that SupCon promotes more compact and stable representation geometry under accent variability. Overall, SupCon provides an effective and model-agnostic regularization strategy for improving accent robustness.

Abstract

Demographic Bias in ASR

Explore similar work

Contrastive Regularization for Accent-Robust ASR

When Entanglement Lower-Bounds Disparity: Auditing and Repairing Demographic Fairness in Audio Understanding Models

Beyond Word Error Rate: Auditing the Diversity Tax in Speech Recognition through Dataset Cartography