Paper deep dive
Do LLM Decoders Listen Fairly? Benchmarking How Language Model Priors Shape Bias in Speech Recognition
Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers, Srinivasan Parthasarathy
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 4/26/2026, 7:58:05 PM
Summary
This research evaluates how different architectural generations of Automatic Speech Recognition (ASR) modelsâspecifically the integration of Large Language Model (LLM) decodersâimpact demographic fairness and robustness. By benchmarking nine models (including Wav2Vec2, Whisper, Qwen3, and Granite-Speech) across five demographic axes and 12 acoustic degradation conditions, the study finds that while LLM decoders do not amplify racial bias, they can introduce pathological hallucinations (especially in Whisper on Indian-accented speech). A key finding is that audio compression levels and encoder design, rather than LLM scale, are the primary drivers of accent fairness and acoustic robustness.
Entities (9)
Relation Signals (4)
Whisper â exhibitshallucinationon â Indian-accented speech
confidence 100% · Whisperâs implicit-LM decoder exhibits pathological hallucination on Indian-accented speech with a non-monotonic insertion-rate spike to 9.62% at large-v3
Silence Injection â amplifiesbiasin â Whisper
confidence 95% · silence injection amplifies Whisper's accent bias up to 4.64x by triggering demographic-selective hallucination
Granite-8B â hasbestethnicityfairness â Ethnicity
confidence 95% · Granite-8B has the best ethnicity fairness, max/min WER = 2.28
Audio Compression â predicts â Accent Fairness
confidence 90% · audio compression predicts accent fairness more than LLM scale
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:As pretrained large language models replace task-specific decoders in speech recognition, a critical question arises: do their text-derived priors make recognition fairer or more biased across demographic groups? We evaluate nine models spanning three architectural generations (CTC with no language model, encoder-decoder with an implicit LM, and LLM-based with an explicit pretrained decoder) on about 43,000 utterances across five demographic axes (ethnicity, accent, gender, age, first language) using Common Voice 24 and Meta's Fair-Speech, a controlled-prompt dataset that eliminates vocabulary confounds. On clean audio, three findings challenge assumptions: LLM decoders do not amplify racial bias (Granite-8B has the best ethnicity fairness, max/min WER = 2.28); Whisper exhibits pathological hallucination on Indian-accented speech with a non-monotonic insertion-rate spike to 9.62% at large-v3; and audio compression predicts accent fairness more than LLM scale. We then stress-test these findings under 12 acoustic degradation conditions (noise, reverberation, silence injection, chunk masking) across both datasets, totaling 216 inference runs. Severe degradation paradoxically compresses fairness gaps as all groups converge to high WER, but silence injection amplifies Whisper's accent bias up to 4.64x by triggering demographic-selective hallucination. Under masking, Whisper enters catastrophic repetition loops (86% of 51,797 insertions) while explicit-LLM decoders produce 38x fewer insertions with near-zero repetition; high-compression audio encoding (Q-former) reintroduces repetition pathology even in LLM decoders. These results suggest that audio encoder design, not LLM scaling, is the primary lever for equitable and robust speech recognition.
Tags
Links
- Source: https://arxiv.org/abs/2604.21276v1
- Canonical: https://arxiv.org/abs/2604.21276v1
Trouble viewing inline? Open PDF directly â
Full Text
64,543 characters extracted from source content.
Expand or collapse full text
Do LLM Decoders Listen Fairly? Benchmarking How Language Model Priors Shape Bias in Speech Recognition Srishti Ginjala 1 , Eric Fosler-Lussier 1 , Christopher W. Myers 2 , Srinivasan Parthasarathy 1 1 The Ohio State University, Columbus, OH, USA 2 Air Force Research Laboratory, USA Abstract As pretrained large language models replace task-specific decoders in speech recognition, a critical question arises: do their text-derived pri- ors make recognition fairer or more biased across demographic groups? We evaluate nine models spanning three architectural generations (CTC with no language model, encoder-decoder with an implicit LM, and LLM- based with an explicit pretrained decoder) onâŒ43,000 utterances across five demographic axes (ethnicity, accent, gender, age, first language) using Common Voice 24 and Metaâs Fair-Speech, a controlled-prompt dataset that eliminates vocabulary confounds. On clean audio, three findings chal- lenge assumptions: LLM decoders do not amplify racial bias, as the most ethnicity-fair competitive model (Granite-8B, max/min WER ratio = 2.28) uses an explicit LLM decoder; Whisperâs implicit-LM decoder exhibits pathological hallucination on Indian-accented speech with a non-monotonic insertion-rate spike to 9.62% at large-v3; and audio compression predicts accent fairness more than LLM scale. We then stress-test these findings under 12 acoustic degradation conditions (noise, reverberation, silence injection, chunk masking) across both datasets, totaling 216 inference runs. Severe degradation paradoxically compresses fairness gaps as all groups con- verge to high WER, but silence injection amplifies Whisperâs accent bias up to 4.64Ăby triggering demographic-selective hallucination. Under mask- ing, Whisper enters catastrophic repetition loops (86% of 51,797 insertions) while explicit-LLM decoders produce 38Ăfewer insertions with near-zero repetition; high-compression audio encoding (Q-former) reintroduces repe- tition pathology even in LLM decoders. These results suggest that audio encoder design, not LLM scaling, is the primary lever for equitable and robust speech recognition. 1 Introduction Language models are increasingly forming the decoder backbone of speech recognition systems. Three architectural generations now coexist: CTC-based systems that use no lan- guage model, mapping audio directly to characters; encoder-decoder systems like Whisper that learn an implicit language model from their training transcriptions; and the newest generation, which routes audio embeddings into a pretrained LLM whose distributional priors are inferred from text-only corpora. Each generation integrates language priors more deeply into recognition. But does a stronger language prior help all speakers equally, or does it encode the biases of its predominantly standard-English training text? ASR is known to exhibit demographic biases: Koenecke et al. (2020) showed that five commercial systems produced roughly double the error rate for Black speakers compared to White speakers, and Tatman (2017) documented gender and dialect disparities. These findings predate the LLM-decoder era. A natural concern follows: pretrained LLMs, whose priors reflect standard written English, may amplify existing gaps when decoding non- 1 arXiv:2604.21276v1 [cs.CL] 23 Apr 2026 Table 1: The nine ASR models evaluated, grouped by the role of the language model in decoding. Audio compression refers to how the raw waveform is transformed before reaching the decoder. Overall WER (%) on each evaluation corpus is shown at right. #ModelGenArchitectureParamsAudio CompressionLM TypeLSCVFS 1Wav2Vec2-large1CTC encoder317MNone (raw frames)No LM1.7922.7232.15 2Whisper-small2Enc-Dec Transformer244MLog-mel 80dImplicit LM3.5016.5911.51 3Whisper-medium2Enc-Dec Transformer764MLog-mel 80dImplicit LM2.9912.598.75 4Whisper-large-v32Enc-Dec Transformer1.5BLog-mel 128dImplicit LM1.9210.967.79 5Qwen3-ASR-0.6B3Audio enc + Qwen30.6BLow (direct)Explicit LLM2.1310.085.89 6Qwen3-ASR-1.7B3Audio enc + Qwen31.7BLow (direct)Explicit LLM1.607.764.73 7Canary-Qwen-2.5B3FastConformer + Qwen-2.5B2.5BMediumExplicit LLM1.617.726.60 8Granite-Speech-2B3Conformer + Q-former + LLM2.0BHigh (Q-former)Explicit LLM1.5310.098.99 9Granite-Speech-8B3Conformer + Q-former + LLM8.0BHigh (Q-former)Explicit LLM2.4210.868.04 standard or accented speech. Moreover, real-world speech is rarely clean: noise, rever- beration, and signal dropout force the decoder to compensate, moving it away from its language prior, potentially amplifying biases invisible on clean benchmarks. This concern is urgent because LLM-based ASR is deployed at scale. Yet, no study has systematically tested whether adding a pretrained language model to the decoder makes recognition more or less equitable, either on clean speech or under acoustic degradation. To our knowledge, we present the first study isolating how the degree of language model integration affects ASR fairness on clean speech and under controlled degradation. We evaluate nine models (one CTC system with no LM, three encoder-decoder systems with an implicit LM, and five LLM-decoder systems with an explicit LM) onâŒ43,000 utterances from Common Voice 24 (Ardila et al., 2020) and Metaâs Fair-Speech (Veliche et al., 2024) across five demographic axes. We test all models under 12 perturbation conditions (noise, reverb, silence, masking) on both datasets, totaling 216 inference runs. Our contributions: (1) the first systematic study of LLM decodersâ effect on ASR fairness across five demographic axes and 12 degradation conditions; (2) evidence that LLM decoders do not amplify racial bias (§3.1); (3) identification of Whisperâs pathological hallucination on Indian-accented speech (9.62% insertion rate) while all LLM decoders remain below 3.1% (§3.2); (4) evidence that audio compression predicts accent fairness more than LLM scale (§3.2); (5) the finding that severe degradation paradoxically compresses fairness gaps, with one critical exception: silence amplifies Whisperâs accent bias up to 4.64Ă(§4); and (6) demonstration that hallu- cination type under degradation is architecture-dependent, with high audio compression reintroducing pathological repetition even in LLM decoders (§4.3). 2 Benchmarking Setup 2.1 Models: three generations of language model integration Table 1 summarizes the nine models in our study. Generation 1: Wav2Vec2-large (Baevski et al., 2020), a CTC encoder fine-tuned on LibriSpeech 960h, maps audio frames to character probabilities with no autoregressive decoder and no language model. Generation 2: three Whisper checkpoints (Radford et al., 2023) (small 244M, medium 764M, large-v3 1.5B), whose autoregressive decoder learns an implicit language model from 680,000 hours of paired transcriptions. Generation 3: five models that route audio embeddings into a pretrained LLM backbone: Qwen3-ASR (0.6B, 1.7B) with low-compression direct audio-token projection (Shi et al., 2026); Canary-Qwen-2.5B with medium compression via FastConformer (NVIDIA, 2025); and Granite-Speech (2B, 8B) with high compression through a Q-former bottleneck sharing one encoder across both sizes (Saon et al., 2025). This yields controlled comparisons: Whisper smallâmediumâlarge-v3 (implicit-LM scaling); Qwen3 0.6Bâ1.7B (explicit-LLM, low compression); Granite 2Bâ8B (explicit-LLM, constant high compression); and Qwen3- 1.7B vs. Granite-2B (âŒsame parameters, different compression). We selected these nine open-weight models from HuggingFace as representative of each architectural generation based on Open ASR Leaderboard performance (as of February 1,2026) and community adoption, subject to our computational budget. 2 2.2 Datasets We evaluate on three English speech corpora. Common Voice 24 test split (Ardila et al., 2020) providesâŒ16,400 crowd-sourced read-speech utterances with self-reported accent (6 groups,n â„50), gender, and age (5 bins); its large Indian-accent subgroup (n=511) enables fine-grained hallucination analysis. Fair-Speech (Veliche et al., 2024) contains âŒ26,470 utterances from U.S. English speakers who all read the same prompted sentences, eliminating vocabulary confounds; any WER differences are attributable to acoustic or model factors. It provides ethnicity (7 groups), gender, age, first language (30+ L1s), and socioeconomic status (3 levels). To our knowledge, we report the first comprehensive multi- model benchmark on Fair-Speech. LibriSpeech test-clean (Panayotov et al., 2015) (2,620 utterances, no demographics) serves as a reference baseline. 2.3 Metrics We report word error rate (WER) as the primary measure of accuracy and max/min ra- tio (MMR) as the fairness metric:M MR= max gâG WER g / min gâG WER g over groups withn â„50. MMR = 1.0 indicates perfect parity; we prefer it over absolute gap because it is scale-invariant. All group-level WERs include 95% bootstrap CIs (200 resamples). Following Morris et al. (2004), we decompose errors into substitutions, insertions, and deletions to distinguish acoustic confusion from pathological decoder-driven errors, such as hallucinations. For perturbation experiments, we additionally report the fairness gap amplification ratio: α= M MR pert /M MR clean , whereα>1 indicates degradation widens demographic gaps andα<1 indicates compression. We classify insertions into repetition (autoregressive loops), syntactic (function words), and content (semantic fabrications) to characterize architecture-specific hallucination behavior under degradation. 2.4 Perturbation Design Choices Real-world audio is rarely clean; degraded signals force the decoder to compensate from its language prior, potentially revealing demographic biases invisible on clean benchmarks. We test all nine models under four perturbation types at three severity levels: additive noise (SNR 20/10/0 dB; source: MUSAN noise corpus (Snyder et al., 2015)); reverberation (RT60 0.3/0.6/1.0 s; source: OpenSLR RIRs (Ko et al., 2017)); silence injection (25/50/75% of utterance duration inserted at a random position); and chunk masking (10/20/30% of audio zeroed out across multiple chunks). Noise and reverb degrade signal quality while preserving temporal structure; silence and masking remove acoustic information entirely, maximally forcing decoder reliance on its language prior. We deliberately exclude standard non-destructive augmentations (e.g., pitch shifting, speed scaling, filtering) because compre- hensive robustness to these has already been studied by Shah et al. (2025). Moreover, unlike signal dropout, these continuous transformations do not differentially stress the decoderâs language model prior, which is the core focus of our analysis. Perturbed audio is generated offline; inference configuration matches the clean evaluation exactly (Appendix H). In total 9 modelsĂ 12 conditionsĂ 2 datasets = 216 inference runs. 3 Clean evaluation 3.1 Ethnicity fairness Baseline performance: All 9 models achieve 1.5-3.5% WER on LibriSpeech test-clean (Table 1), confirming base competence. Performance diverges on diverse speech: on Fair- Speech, Gen 3 models lead (Qwen3-1.7B: 4.73%), Gen 2 models range from 7.8-11.5%, and Wav2Vec2-large reaches 32.15%, an 18-fold increase over its LibriSpeech WER reflecting a domain mismatch between its training data and Fair-Speechâs mobile-recorded speech. WER by Ethnicity: Figure 1a presents WER by ethnicity on Fair-Speech (full table in Appendix A). The most consistent finding in this study is that Black/African-American 3 Asian Pacific Islander Hispanic Native American Middle Eastern White Black/A Ethnicity W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Model 28.730.430.227.332.526.341.6 4.66.47.69.79.510.518.6 3.84.16.27.57.69.812.4 3.33.75.67.46.19.410.2 3.54.43.84.15.83.011.0 2.83.52.93.64.42.88.5 3.44.73.84.75.84.911.7 5.35.45.66.26.310.013.4 6.15.35.55.65.67.612.1 WER (%) by Ethnicity (Fair-Speech) 5 10 15 20 25 30 35 40 WER (%) (a) Ethnicity WER on Fair-Speech. usenglandcanadaaustraliaindianafrican Accent W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Model 16.416.511.813.225.223.3 11.513.16.38.017.620.3 8.89.44.26.413.211.5 7.57.73.65.219.013.0 7.58.14.15.311.19.6 5.86.74.64.47.88.0 5.86.34.84.07.28.0 8.39.03.15.010.410.0 10.69.44.34.712.916.1 WER (%) by Accent (Common Voice) 5.0 7.5 10.0 12.5 15.0 17.5 20.0 22.5 25.0 WER (%) (b) Accent WER on Common Voice 24. Figure 1: WER (%) by (a) ethnicity and (b) accent. Black/A speakers face the highest WER everywhere. Indian/African accents are hardest, with Whisper-large-v3 underperforming small on Indian accent due to hallucination. speakers have the highest WER across all nine models, regardless of generation or architec- ture. This holds for Gen 1 (Wav2Vec2: 41.6%), Gen 2 (Whisper-large-v3: 10.2%), and Gen 3 (Qwen3-1.7B: 8.5%). The gap is statistically significant for 8 of 9 models: bootstrap 95% CIs for Black/A and White WER do not overlap (Table 16). The sole exception is Whisper- large-v3, where Black/A 10.25% [9.61, 10.81] and White 9.42% [8.21, 10.64] overlap, where Whisperâs scaling genuinely narrows the ethnicity gap, even as it simultaneously worsens accent fairness through hallucination (§3.2). This underscores a challenge with intersectional fairness where improvements on one sensitive axis often degrades another. Pre-trained LLM priors do not widen racial gaps. Contrary to the above hypothesis, Gen 3 models do not exhibit worse ethnicity fairness than Gen 2. Among competitive models, Granite-8B (Gen 3, MMR = 2.28) achieves the best ethnicity fairness, outperforming all three Whisper models (MMR 3.13-4.04; Table 4). Gender disparities are negligible on Common Voice (max MMR = 1.12) but substantial on Fair-Speech (Qwen3-0.6B gender MMR = 2.42), likely due to Fair-Speechâs controlled prompts revealing acoustic-level effects masked by Common Voiceâs heterogeneous sentences (Appendix C). A measurement paradox emerges: Qwen3-1.7B achieves the lowest absolute Black/A WER (8.45%) yet the highest relative gap (+203%) because its White WER is extremely low (2.79%). Conversely, Wav2Vec2 achieves the smallest relative gap (MMR = 1.58), but this is merely âlow-accuracy parityâ where all groups suffer unacceptably high error rates (e.g., 41.6% Black/A). A recommendation here is that practitioners should report both absolute WER per group and relative disparity metrics. 3.2 Accent fairness and decoder hallucination Baseline performance: Figure 1b presents WER across six accent groups on Common Voice 24 (full table in Appendix A). Indian and African accents are hardest for every model, but error sources differ dramatically by architecture. Qwen3-1.7B is the most equi- table (accent MMR = 1.82, range 4.41-8.03%), while Whisper-large-v3 is the least equitable (MMR = 5.34) despite its best overall WER. The surprising paradox: Whisper-large-v3âs Indian- accent WER (19.0%) is worse than Whisper-smallâs (17.6%). Compression predicts accent fairness. Among Gen 3 models at comparable scale (â2B), low-compression Qwen3-1.7B (accent MMR = 1.82) outperforms medium-compression Canary-2.5B (2.00) and high-compression Granite-2B (3.30). This dynamically shifts for ethnicity: high-compression Granite-2B (2.53) outperforms low-compression Qwen3-1.7B (3.03), suggesting accent depends on fine-grained phonetic and prosodic cues that are easily lost in compression while ethnicity-relevant patterns may be recovered despite compression. Whisper hallucinates on Indian-accented speech. Table 2 shows that Whisperâs insertion rate on Indian accent (n=511) follows a non-monotonic trajectory: small 3.22%âmedium 1.53%â large-v3 9.62%. At large-v3, insertions become the dominant error type (50.7% of 4 Table 2: Whisper scaling on Indian-accented speech (Common Voice 24,n=511). Insertion rate is non-monotonic: medium improves over small, then large-v3 spikes pathologically. ModelParamsIndian WERIns. RateSub. RateIns. % of Errors Whisper-small244M17.6%3.22%12.92%18.3% Whisper-medium764M13.2%1.53%9.99%11.6% Whisper-large-v31.5B19.0%9.62%8.32%50.7% wav2vec2-large whisper-small whisper-medium whisper-large-v3 qwen3-asr-0.6bqwen3-asr-1.7b canary-qwen-2.5b granite-speech-3.3-2bgranite-speech-3.3-8b 0 20 40 60 80 100 % of Insertions Hallucination Category Distribution (CV) Syntactic Content Repetition Figure 2: Hallucination categories on Common Voice 24. Whisper-large-v3 is driven by repetition loops. Graniteâs repetition pathology is diagnosed under signal dropout. all errors). All five Gen 3 models remain below 3.1% on the same utterances (Appendix G). This pathology is accent-selective: Whisper-large-v3âs insertion rates on US (0.96%) and Canadian (0.52%) accents are comparable to other models. At the 9.62% insertion rate from Table 2, Whisper-large-v3 produces 495 insertions on the 511 Indian-accent utterances. Table 13 (Appendix G) decomposes these by type: 43% are repetition loops, 48% syntactic completions, and 9.3% content hallucinations, semantic fabrications that alter meaning (e.g., âgrow sea monkeysâââgo and see what happensâ; Table 3). Figure 2 shows the all-accent category distribution. Gen 3 modelsâ far rarer insertions are predominantly benign function words (Table 13). 3.3 Scaling trajectories Scaling effects. Scaling effects differ by architecture (Figure 3). Qwen3 (0.6Bâ1.7B, low compression) improves both accuracy and fairness: accent MMR drops from 2.72 to 1.82 and ethnicity MMR from 3.68 to 3.03. Whisper scaling improves ethnicity fairness (MMR: 4.04â3.28â3.13) but worsens accent fairness (MMR: 3.23â3.15â5.34) due to the Indian- accent hallucination pathology (§3.2). Granite (2Bâ8B, high compression) shows dataset- dependent effects: ethnicity fairness improves on Fair-Speech (MMR: 2.53â2.28) but accent fairness worsens on Common Voice (MMR: 3.30â3.74), suggesting the Q-former bottleneck limits the benefit of additional LLM capacity for acoustically diverse speech. Table 3: Hallucinations on Indian-accented speech. Whisper-large-v3 fabricates semantic content; Canary-Qwen-2.5B inserts function words. Insertions bolded. ModelReferenceHypothesisCategory Whisper-large-v3perhaps you should just grow sea monkeys perhaps we should just go and see what happens Content Whisper-large-v3playing the 3 year old Kimberley Macdonald playing the 3 year old kim ming mc- donnell Content Whisper-large-v3 boyd was not always popular with rank and file officers the 1st day of the 1st day of the 1st day of. . . Repetition Canary-Qwen-2.5Bin korea carnations express admira- tion love and gratitude in korea the carnations express ad- miration love and gratitude Syntactic Canary-Qwen-2.5B he is noted for his studies of plants in the south of africa he is noted for his studies of plants in in the south of africa Syntactic 5 1.0B Parameters (log scale) 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 Max/Min Ratio (MMR) Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Gr-2B Gr-8B Common Voice: Accent MMR Whisper Qwen3 Granite 1.0B Parameters (log scale) 2.25 2.50 2.75 3.00 3.25 3.50 3.75 4.00 Max/Min Ratio (MMR) Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Gr-2B Gr-8B Fair-Speech: Ethnicity MMR Whisper Qwen3 Granite Scaling x Fairness: Three Architecture Families Figure 3: Scaling trajectories. Qwen3 improves both accuracy and fairness. Whisper worsens accent fairness via hallucination. Granite shows dataset-dependent effects. Additional demographic analyses (first language, age, and socioeconomic status) appear in Appendices F-E. Key findings: Whisperâs implicit LM creates extreme L1 sensitivity (L1 MMR up to 15.05; Gen 3â€5.38); ageĂarchitecture interactions flip which group is worst; SES gaps are moderate (MMR 1.13-2.17). 4 Fairness under acoustic degradation Having established clean-audio fairness patterns, we now test whether they hold when acoustic degradation forces the decoder to increasingly rely on its language prior. 4.1 Architecture-specific robustness Degradation robustness across ASR generations. The three ASR generations differ markedly in degradation robustness (Figure 4). Under 30% chunk masking on Fair-Speech, Wav2Vec2-large (no LM) reaches 58.9% WER, nearly double its clean 32.1%. Among Gen 3 models sharing Qwen LLM backbones, compression level predicts the degradation penalty: low-compression Qwen3-1.7B degrades to 33.9%, while medium-compression Canary de- grades worse to 35.8%-a consistent trend observed across all four perturbation types. Whis- per spans a wide range: small (57.8%) degrades comparably to Wav2Vec2, while large- v3 (36.8%) approaches Gen 3 performance. Highly-compressed Granite-2B (43.4%) and Granite-8B (38.4%) degrade much closer to Whisper. This confirms that audio bottleneck compression-not solely the LLM text pretraining data-dictates acoustic robustness. Impact of perturbation type on robustness. Reverberation is the mildest perturbation for models with language priors: at RT60 = 1.0 s, Qwen3-1.7B barely increases (4.7%â5.0%) and Whisper-large-v3 rises to only 8.4%, while Wav2Vec2 reaches 35.0%. Additive noise at SNR = 0 dB falls between, with Qwen3-1.7B (7.4%) and Whisper-large-v3 (12.1%) demon- strating that stronger language priors better compensate for noise. Under silence injection, WER degrades minimally for Qwen3 and Canary, but increases almost linearly with silence duration for Whisper and Granite. These patterns hold on Common Voice (Figure 17 in Appendix J): under 30% masking, Qwen3-1.7B (35.7%) outperforms all Whisper checkpoints (42.4-63.0%) and Wav2Vec2 (57.6%), while silence injection uniquely triggers demographic- selective hallucination (§4.2). Clean20dB10dB0dB 10 20 30 40 50 WER (%) Additive Noise Clean0.3s0.6s1.0s 5 10 15 20 25 30 35 Reverberation Clean25%50%75% 5 10 15 20 25 30 35 Silence Injection Clean10%20%30% 10 20 30 40 50 60 Chunk Masking WER Degradation Under Perturbation (Fair-Speech) W2V2-LWh-SWh-MWh-LQ3-0.6BQ3-1.7BCan-2.5BGr-2BGr-8B Figure 4: WER degradation curves on Fair-Speech. Masking produces the steepest degrada- tion. Qwen3-1.7B (green) is the most robust model. 6 SNR 20dB SNR 10dB SNR 0dB Rev 0.3s Rev 0.6s Rev 1.0s Sil 25% Sil 50% Sil 75% Mask 10% Mask 20% Mask 30% W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B 1.021.000.851.001.011.020.950.950.950.910.840.79 0.910.750.590.910.870.760.600.750.500.480.390.40 0.890.920.880.991.091.120.980.860.870.620.460.44 1.011.090.971.011.041.130.900.850.990.650.510.44 0.960.920.780.961.000.990.930.900.870.510.400.36 0.960.960.870.970.960.980.920.910.890.580.460.42 0.991.010.770.961.000.940.800.760.760.570.430.38 0.961.100.840.941.011.110.920.900.830.710.570.51 1.021.151.121.011.021.090.920.950.980.770.630.58 NoiseReverbSilence Masking Fairness Gap Amplification Under Perturbation: Ethnicity (Fair-Speech) 0.5 1.0 1.5 α (MMR_pert / MMR_clean) (a) Ethnicity Gap Amplification (Fair-Speech) SNR 20dB SNR 10dB SNR 0dB Rev 0.3s Rev 0.6s Rev 1.0s Sil 25% Sil 50% Sil 75% Mask 10% Mask 20% Mask 30% W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B 0.980.950.770.981.021.040.850.840.790.770.660.58 0.870.760.721.321.110.954.641.101.060.600.560.63 0.860.791.040.850.900.911.121.231.180.650.580.99 0.580.570.460.590.740.700.941.241.380.520.300.32 1.011.010.991.011.081.160.940.750.770.630.570.48 1.361.631.501.381.341.341.181.101.231.220.860.72 1.031.221.391.071.031.010.990.941.100.910.730.75 0.970.820.690.840.860.730.540.660.510.740.420.40 1.060.850.861.061.021.191.020.830.760.890.470.43 NoiseReverbSilence Masking Fairness Gap Amplification Under Perturbation: Accent (Common Voice) 2 4 α (MMR_pert / MMR_clean) (b) Accent Gap Amplification (Common Voice 24) Figure 5: Fairness gap amplification (α) by (a) ethnicity and (b) accent. Blue (α<1): compresses gaps; red (α>1): widens gaps. Masking universally compresses gaps. Noise amplifies Qwen3-1.7B bias (α = 1.63); silence redistributes Whisper-large-v3 bias (§4.2). 4.2 Fairness gap amplification Figure 5 presents the fairness gap amplification ratio (α= M MR pert /M MR clean ) for ethnic- ity on Fair-Speech (panel a) and accent on Common Voice (panel b) across all nine models and 12 perturbation conditions. Severe masking compresses fairness gaps. Under 30% masking on Fair-Speech, every modelâs ethnicity MMR compresses (α= 0.36-0.79; clean MMR 1.58-4.04, Table 4). This mirrors the âlow-accuracy parityâ from §3.1: when all groups suffer similarly catastrophic error rates, relative gaps shrink. Accent fairness on Common Voice compresses similarly under masking (α = 0.32-0.99; Figure 5b). At harshest noise (SNR = 0 dB), most models also compress ethnicity gaps (α= 0.59-0.97), though Granite-8B amplifies (α= 1.12). Reverb at RT60 = 1.0 s produces mild ethnicity amplification for Whisper-large-v3 (α= 1.13), Whisper- medium (1.12), Granite-2B (1.11), and Granite-8B (1.09), remaining close to unity. Accent gaps amplify under noise and silence. On Common Voice accent fairness (Fig- ure 5b), amplification patterns are more prominent. Qwen3-1.7B, the most equitable model on clean audio (accent MMR = 1.82), shows the strongest amplification:α= 1.63 at SNR = 10 dB (MMR: 1.82â2.96), with Indian-accent WER rising from 7.8% to 11.3% while Canadian barely changes (4.6%â3.9%; Table 18). Canary-2.5B amplifies similarly un- der noise (α= 1.39 at SNR = 0 dB). Equitable clean performance does not guarantee equitable degradation. Whisper-small exhibits the most extreme amplification:α= 4.64 at 25% silence (MMR: 3.23â15.0), driven by African-accent WER increasing 4.2Ăwhile Canadian barely changes. This effect is non-monotonic, collapsing toαâ1.1 at 50-75% silence, suggesting a threshold where mild silence triggers catastrophic hallucination loops for specific accent groups. Whisper-large-v3 under 75% silence also amplifies accent bias (α= 1.38; MMR: 5.38â7.45), but the mechanism is unexpected: Indian-accent WER decreases from 19.2% to 14.5% as silence disrupts the hallucination pathway from §3.2, while England-accent WER quadruples from 7.7% to 30.4%. Perturbation redistributes decoder bias across demographic groups rather than uniformly amplifying it. 7 4.3 Hallucination under signal dropout Both masking and silence injection, limits the reliance on acoustic information, forcing the decoder to generate text from its prior alone. The resulting hallucination patterns reveal fundamental architectural differences. Under 30% masking on Fair-Speech, Whisper-small produces 51,797 insertions, of which 86% are repetition loops: catastrophic autoregressive cycling where the decoder repeats phrases hundreds of times. Whisper-large-v3 is more contained (10,365 insertions, 35% repetition) but still repetition-prone. By contrast, Qwen3- 1.7B produces only 1,353 insertions (38Ăfewer than Whisper-small) with just 3% repetition; its rare insertions are primarily syntactic completions (54%) and content substitutions (44%). Wav2Vec2 (no LM) produces content-classified insertions (56%, phonetic garble) with<2% repetition, strongly suggesting that repetition loops are a decoder-driven phenomenon. Compression reintroduces repetition.Among Gen 3 models, audio compression predicts hallucination type under degradation (Figure 6). Qwen3 (low compression) maintains<4% repetition across all conditions. But Granite-Speech (high compression, Q-former bottleneck) exhibits 49-57% repetition under 30% masking, closer to Whisper than to Qwen, with 7,689 (2B) and 10,187 (8B) total insertions. Canary (medium compression) falls between at 30% repetition and 4,552 insertions. Just as compression predicts accent fairness on clean audio (§3.2), it also predicts hallucination behavior under degradation: high audio compression induces Whisper-like repetition pathology even when the decoder is an explicit LLM. Silence perturbations show similar patterns: Whisper-small produces 57-60% repetition, Qwen3 <4%, and Granite 31-43% (full breakdowns in Appendix K). W2V2-LWh-SWh-MWh-LQ3-0.6BQ3-1.7BCan-2.5BGr-2BGr-8B 0 20 40 60 80 100 Insertion Type Distribution (%) n=5,896n=51,797n=16,474n=10,365n=1,756n=1,353n=4,552n=7,689n=10,187 Hallucination Types Under 30% Masking (Fair-Speech) Repetition Syntactic (function words) Content (semantic) Figure 6: Hallucination type distribution under masking.ndenotes insertion count. Whisper-small is dominated by repetition loops; Qwen3 shows near-zero repetition; Granite reintroduces repetition due to high compression. 5 Discussion 5.1 Accuracy does not imply fairness; degradation does not imply inequity Figure 7 synthesizes the accuracy-fairness tradeoff. On clean speech (panel a), Qwen3-1.7B achieves the best WER (4.73%) with competitive ethnicity fairness (MMR = 3.03), while Granite-8B achieves the best ethnicity fairness among competitive models (MMR = 2.28) at higher WER (8.04%). On the accent axis, Qwen3-1.7B also achieves the best MMR (1.82), making it the strongest overall candidate for fair deployment. Under degradation (panel b), all models shift rightward (higher WER) and downward (lower MMR), converging toward a degenerate frontier where low relative disparity coexists with poor absolute performance, echoing the âlow-accuracy parityâ paradox from §3.1. Practical recommendations. Qwen3-1.7B dominates on accuracy, accent fairness, and robustness; Granite-8B is preferable for ethnicity fairness. Whisper-large-v3 should not be deployed on accented speech without hallucination mitigation (Baranski et al., 2025), 8 51015202530 Overall WER (%) 1.5 2.0 2.5 3.0 3.5 4.0 Ethnicity MMR (higher = less fair) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Accuracy vs. Fairness (Fair-Speech) Gen 1 (CTC) Gen 2 (Enc-Dec) Gen 3 (LLM-ASR) (a) Clean audio. 1020304050 WER (%) 1.5 2.0 2.5 3.0 3.5 4.0 MMR (Ethnicity) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Accuracy vs. Fairness Under Degradation (Fair-Speech) Condition Clean SNR 10dB Mask 20% (b) Under degradation (SNR 10 dB and 20% mask- ing). Figure 7: Accuracy vs. ethnicity fairness. (a) Clean: Qwen3-1.7B and Granite-8B define Pareto frontier. (b) Under degradation: models converge to degenerate frontier where low disparity coexists with poor performance. Arrows trace clean to 20% masking. especially with silence or signal dropout. Low-compression audio encoding (as in Qwen3) provides robustness against both accuracy loss and bias amplification. 5.2 Limitations Our study is limited to English read and prompted speech and may not generalize to multi- lingual or spontaneous ASR. Training data is a confound: we cannot fully disentangle the decoderâs contribution from training corpora differences. Granite-8Bâs higher LibriSpeech WER than Granite-2B is likely a verbose generation artifact and does not affect fairness comparisons. Perturbation conditions are synthetic and applied in isolation. Common Voice sentences are drawn from Wikipedia, and LibriSpeech from Project Gutenberg, both standard components of LLM pretraining corpora. Gen 3 modelsâ LLM decoders may have encountered exact transcripts during pretraining. Two observations constrain this concern: if contamination drove improvements, we would expect uniform WER gains across groups, yet Black/A gaps persist (+203% for Qwen3-1.7B); and Gen 3 models maintain their advantage even under severe masking that removes intelligible acoustic content. In such conditions, memorized text cannot aid decoding, suggesting that the observed gains are unlikely to be driven by test-set contamination 6 Related work ASR demographic bias is extensively documented across ethnicity (Koenecke et al., 2020), gender and dialect (Tatman, 2017), and Indian-accented English (Javed et al., 2023). While recent multi-axis benchmarks and controlled datasets like Fair-Speech offer standardized evaluation (Rai et al., 2025; Veliche et al., 2024), they primarily assess commercial systems. Unlike these studies, which evaluate limited systems on isolated axes, we specifically investigate how decoder architecture and language model integration drive these disparities. On hallucination, Koenecke et al. (2024) foundâŒ1% of Whisper outputs hallucinate (38% harmful); Baranski et al. (2025) showed non-speech audio triggers hallucinations and proposed energy-based filtering; Frieske & Shi (2024) developed taxonomy-based detection; and Atwany et al. (2025) traced hallucinations to distribution shifts. (Koudounas et al., 2025) introduced a hallucination benchmark with a fine-grained taxonomy across architectures and acoustic conditions, but none of these studies connect hallucination to demographic bias. Robustness studies further highlight performance drops under noise and domain shifts (Shah et al., 2025; Kumar et al., 2025), yet demographic breakdowns in these contexts remain sparse. Wei et al. (2026) examine bias in spoken QA for multimodal LLMs but do not study transcription quality or decoder architecture. No prior study examines whether adding a pretrained LLM decoder systematically affects fairness under acoustic degradation. 9 7 Conclusion We studied how three generations of language model integration affect demographic fairness across nine models, five demographic axes, and 12 degradation conditions. LLM decoders do not amplify racial bias, but stronger priors inflate relative disparity as majority-group WER approaches zero. Severe degradation generally compresses fairness gaps, with one crit- ical exception: silence amplifies Whisper âs accent bias (α= 4.64) via demographic-selective hallucination. Under masking, Whisper produces catastrophic repetition loops (86% of insertions) while explicit-LLM decoders produce 38Ăfewer with near-zero repetition; high audio compression reintroduces this pathology, identifying audio encoder architecture as the primary lever for equitable recognition. Reproducibility Statement To facilitate reproducibility, all model weights and their HuggingFace identifiers, along- side our exact inference parameters and text normalization steps, are fully reported in Appendix H. The Fair-Speech, Common Voice 24, MUSAN, and OpenSLR RIR datasets are publicly available. All code for our data preprocessing, perturbation generation, and evaluation pipeline will be open-sourced upon publication. Ethics Statement This study evaluates ASR systems for demographic bias to inform fairer deployment. All demographic labels are self-reported; we report group-level statistics only and no individual speakers can be identified. Our finding that Black/African-American speakers experience the highest WER across all nine models should motivate targeted improvement, not rationalize exclusion from ASR-dependent services. Hallucination examples (Table 3) illustrate failure modes from public datasets. Our fairness metrics are imperfect proxies for harm; a low MMR can mask uniformly poor performance (§3.1). Practitioners should evaluate deployment-specific impacts beyond aggregate metrics. References Rosana Ardila, Megan Branez, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. Common Voice: A massively-multilingual speech corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, p. 4218â4222, 2020. Hanin Atwany, Abdul Waheed, Rita Singh, Monojit Choudhury, and Bhiksha Raj. Lost in transcription, found in distribution shift: Demystifying hallucination in speech foundation models. In Findings of the Association for Computational Linguistics (ACL), p. 23181â23203, 2025. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, p. 12449â12460, 2020. Marek Baranski et al. Whisper hallucinations: Evaluating and mitigating the generation of plausible-sounding but incorrect transcriptions. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025. Rita Frieske and Bertram E. Shi. Hallucinations in neural Automatic Speech Recognition: Identifying errors and hallucinatory models. arXiv preprint arXiv:2401.01572, 2024. Tahir Javed et al. Svarah: Evaluating English ASR systems on Indian accents. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. Tom Ko, Vijayaditya Peddinti, Daniel Povey, and Sanjeev Khudanpur. A study on data augmentation of reverberant speech for robust speech recognition. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 5220â5224. IEEE, 2017. 10 Allison Koenecke, Andrew Nam, Emily Lake, Joe Nudell, Minber Quartey, Zion Mengesha, Connor Tobin, Drew R Harris, Howard Vaisey, and Alexander Hogan. Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences, 117(14): 7684â7689, 2020. Allison Koenecke, Anna Choi, Katelyn Mei, Hilke Schellmann, and Mona Sloane. Careless Whisper: Speech-to-text hallucination harms. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, p. 1672â1681. ACM, 2024. Alkis Koudounas, Moreno La Quatra, Manuel Giollo, Sabato Marco Siniscalchi, and Elena Baralis. SHALLOW: A hallucination benchmark for speech foundation models. arXiv preprint arXiv:2510.16567, 2025. Shashi Kumar, Iuliia Thorbecke, Sergio Burdisso, EsaĂș Villatoro-Tello, K. E. Manjunath, Kadri Hacio Ìglu, Pradeep Rangappa, Petr Motlicek, Aravind Ganapathiraju, and Andreas Stolcke. Performance evaluation of SLAM-ASR: The good, the bad, the ugly, and the way forward. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing Workshops (ICASSPW), p. 1â5. IEEE, 2025. Andrew Cameron Morris, Viktoria Maier, and Phil Green. From WER and RIL to MER and WIL: Improved evaluation measures for connected speech recognition. Proceedings of Interspeech, 2004. NVIDIA. Canary-Qwen-2.5B: A speech-augmented language model for multilingual ASR, 2025. Available at: https://huggingface.co/nvidia/canary-qwen-2.5b. Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. LibriSpeech: An ASR corpus based on public domain audio books. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 5206â5210. IEEE, 2015. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning (ICML), p. 28492â28518. PMLR, 2023. Sarmila Rai et al. ASR-FAIRBENCH: A comprehensive fairness benchmarking framework for Automatic Speech Recognition. In Proceedings of Interspeech, 2025. George Saon et al. Granite-Speech: Open-source speech-aware LLMs with strong English ASR capabilities. arXiv preprint arXiv:2505.08699, 2025. Muhammad A. Shah, David Solans Noguero, Mikko A. HeikkilĂ€, Bhiksha Raj, and Nicolas Kourtellis. Speech Robust Bench: A robustness benchmark for speech recognition. In Proceedings of the International Conference on Learning Representations (ICLR), 2025. Jing Shi et al. Qwen3-ASR technical report: Advancing audio-language understanding. arXiv preprint arXiv:2601.21337, 2026. David Snyder, Guoguo Chen, and Daniel Povey. MUSAN: A music, speech, and noise corpus. arXiv preprint arXiv:1510.08484, 2015. Rachael Tatman. Gender and dialect bias in YouTubeâs automatic captions. In Proceedings of the First ACL Workshop on Ethics in Natural Language Processing, p. 53â59, 2017. Irina-Elena Veliche et al. Fair-Speech: Evaluating fairness in speech foundation models. In Proceedings of Interspeech, 2024. Sheng-Lun Wei, Yu-Ling Liao, Yen-Hua Chang, Hen-Hsen Huang, and Hsin-Hsi Chen. Bias in the ear of the listener: Assessing sensitivity in audio language models across linguistic, demographic, and positional variations. In Findings of the Association for Computational Linguistics (EACL), p. 1570â1589, 2026. 11 Table 4: WER (%) by ethnicity across ASR models on Fair-Speech. ModelWhiteBlack/AAHispanicAsianNative AmericanPacific IslanderMiddle EasternMMRGap (%) wav2vec2-large26.341.630.228.727.330.432.51.5858.4 whisper-small10.518.67.64.69.76.49.54.04304.2 whisper-medium9.812.46.23.87.54.17.63.28228.2 whisper-large-v39.410.25.63.37.43.76.13.13213.2 qwen3-asr-0.6b3.011.03.83.54.14.45.83.68267.5 qwen3-asr-1.7b2.88.52.92.83.63.54.43.03203.3 canary-qwen-2.5b4.911.73.83.44.74.75.83.47246.7 granite-speech-3.3-2b10.013.45.65.36.25.46.32.53153.3 granite-speech-3.3-8b7.612.15.56.15.65.35.62.28127.9 Table 5: WER (%) by accent across ASR models on Common Voice 24. ModelAfricanAustraliaCanadaEnglandIndianUsMMRGap (%) wav2vec2-large23.313.211.816.525.216.42.13113.0 whisper-small20.38.06.313.117.611.53.23222.9 whisper-medium11.56.44.29.413.28.83.15215.3 whisper-large-v313.05.23.67.719.07.55.34434.0 qwen3-asr-0.6b9.65.34.18.111.17.52.72171.6 qwen3-asr-1.7b8.04.44.66.77.85.81.8282.0 canary-qwen-2.5b8.04.04.86.37.25.82.00100.2 granite-speech-3.3-2b10.05.03.19.010.48.33.30230.2 granite-speech-3.3-8b16.14.74.39.412.910.63.74274.5 A Ethnicity and Accent WER tables Table 4 reports WER by ethnicity on Fair-Speech; Table 5 reports WER by accent on Common Voice 24. B Whisper scaling trajectory Figure 8 shows the Whisper three-point scaling trajectory. Accent MMR worsens at large-v3 even as ethnicity MMR improves, driven by the Indian-accent insertion pathology (§3.2). Scaling within Whisper shows architecture-dependent effects: ethnicity fairness improves (MMR: 4.04â3.28â3.13) while accent fairness worsens (MMR: 3.23â3.15â5.34). 0.20.40.60.81.01.21.4 Parameters (B) 3.5 4.0 4.5 5.0 Accent MMR Wh-S Wh-M Wh-L (a) Accent Fairness 0.20.40.60.81.01.21.4 Parameters (B) 3.2 3.4 3.6 3.8 4.0 Ethnicity MMR Wh-S Wh-M Wh-L (b) Ethnicity Fairness 0.20.40.60.81.01.21.4 Parameters (B) 2 3 4 5 6 7 8 9 10 Insertion Rate (%) Wh-S Wh-M Wh-L (c) Indian Accent Insertions Whisper Scaling Trajectory: Fairness Degrades with Scale Figure 8: Whisper three-point scaling trajectory. Accent MMR worsens at large-v3 (5.34) even as ethnicity MMR improves (3.13), driven by the Indian-accent insertion pathology. C Gender WER Table 6 reports WER by gender on Common Voice 24 and Table 7 on Fair-Speech. As noted in §3.1, gender disparities are negligible on Common Voice (max MMR = 1.12) but substantial on Fair-Speech (Qwen3-0.6B gender MMR = 2.42). Males are consistently 1.5- 2.4Ăworse than females across Gen 3 models on Fair-Speech. This discrepancy likely reflects Fair-Speechâs controlled prompts, which reveal acoustic-level gender effects masked by Common Voiceâs heterogeneous sentences. Figures 9 and 10 visualize these differences. 12 Table 6: WER (%) by gender across ASR models on Common Voice 24. ModelFemaleMaleMMRGap (%) wav2vec2-large26.924.51.1010.2 whisper-small18.317.51.054.6 whisper-medium13.713.01.065.6 whisper-large-v311.111.11.000.2 qwen3-asr-0.6b11.010.61.043.8 qwen3-asr-1.7b8.08.01.000.2 canary-qwen-2.5b8.67.71.1212.3 granite-speech-3.3-2b10.710.31.044.5 granite-speech-3.3-8b11.010.71.033.3 Table 7: WER (%) by gender across ASR models on Fair-Speech. ModelfemalemaleMMRGap (%) wav2vec2-large27.537.91.3837.7 whisper-small8.515.31.7979.5 whisper-medium7.410.51.4242.4 whisper-large-v36.89.01.3332.8 qwen3-asr-0.6b3.68.82.42142.4 qwen3-asr-1.7b3.16.82.18118.0 canary-qwen-2.5b4.69.11.9595.0 granite-speech-3.3-2b7.311.21.5353.3 granite-speech-3.3-8b6.410.11.5959.1 femalemale Gender 0 5 10 15 20 25 WER (%) WER by Gender (Common Voice) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Figure 9: WER by gender on Common Voice 24. Gender gaps are minimal across all models (max MMR = 1.12). D Age WER Tables 8 and 9 report WER by age group on Common Voice 24 and Fair-Speech, respectively. A notable ageĂarchitecture interaction emerges on Fair-Speech: the 46-65 age group is the worst for Whisper models (Whisper-medium: 11.8% WER) but the best for Qwen3 models (Qwen3-1.7B: 2.8%). Figure 11 shows the Common Voice results; Figure 12 shows Fair-Speech. E Socioeconomic status WER Table 10 reports WER by self-reported socioeconomic status on Fair-Speech. Medium-SES speakers have the highest WER across 8 of 9 models (e.g., Whisper-small: Medium 13.8% vs. Low 10.6% vs. Affluent 6.4%), with Affluent consistently lowest (SES MMR: 1.13-2.17). Gen 3 models show slightly narrower gaps (MMR 1.67-1.93) than Gen 2 (1.94-2.17). Figure 13 visualizes the gap. F Full first-language WER (21 languages) Whisper âs implicit-LM decoder creates extreme first-language (L1) sensitivity on Fair-Speech. Across 21 individual L1s (Table 11), Whisper-small has an L1 MMR of 15.05 (Filipino: 1.0%, French: 14.3%). Whisper-large-v3 shows a 2.2Ăratio between L1-English (8.2%) and L1-Mandarin (3.8%) speakers. Gen 3 models are substantially more robust: Qwen3-1.7B narrows this to only 1.1Ă(L1-English 5.1% vs. L1-Mandarin 4.6%). The counterintuitive finding that L1-English speakers have the highest WER is an ethnicity confound: L1-English speakers in Fair-Speech are disproportionately Black/A and Native American. Figure 14 visualizes results by L1 group. 13 femalemale Gender 0 5 10 15 20 25 30 35 WER (%) WER by Gender (Fair-Speech) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Figure 10: WER by gender on Fair-Speech. Males are consistently 1.5-2.4Ăworse than females across Gen 3 models, revealed by Fair-Speechâs controlled prompts. Table 8: WER (%) by age across ASR models on Common Voice 24. ModelTeensTwentiesThirtiesFortiesFiftiesMMRGap (%) wav2vec2-large27.228.422.117.620.61.6161.1 whisper-small20.020.415.710.615.31.9292.0 whisper-medium15.415.410.88.412.31.8484.1 whisper-large-v312.313.49.17.49.71.8181.2 qwen3-asr-0.6b12.312.69.47.56.81.8585.1 qwen3-asr-1.7b9.39.56.95.45.21.8282.5 canary-qwen-2.5b10.09.46.85.46.11.8585.4 granite-speech-3.3-2b11.812.08.88.18.21.4848.4 granite-speech-3.3-8b13.712.29.19.19.81.5150.8 Table 9: WER (%) by age across ASR models on Fair-Speech. Model18-2223-3031-4546-65MMRGap (%) wav2vec2-large28.529.737.725.71.4746.9 whisper-small6.88.213.412.81.9898.3 whisper-medium4.87.08.911.82.44144.3 whisper-large-v35.05.87.411.42.25125.1 qwen3-asr-0.6b4.34.98.23.22.60160.1 qwen3-asr-1.7b3.53.96.42.82.30130.5 canary-qwen-2.5b4.84.78.35.81.7776.9 granite-speech-3.3-2b5.36.49.711.42.15115.2 granite-speech-3.3-8b5.36.29.08.91.7170.7 Table 10: WER (%) by socioeconomic status across ASR models on Fair-Speech. ModelLowMediumAffluentMMRGap (%) wav2vec2-large30.934.430.51.1312.8 whisper-small10.613.86.42.17116.7 whisper-medium8.89.34.81.9494.3 whisper-large-v38.18.03.72.17117.0 qwen3-asr-0.6b5.27.24.21.7272.2 qwen3-asr-1.7b4.25.73.41.6767.0 canary-qwen-2.5b6.37.53.91.9392.5 granite-speech-3.3-2b9.19.55.51.7070.5 granite-speech-3.3-8b8.08.65.11.6868.3 14 teenstwentiesthirtiesfortiesfifties Age 0 5 10 15 20 25 WER (%) WER by Age (Common Voice) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Figure 11: WER by age group on Common Voice 24. 18-2223-3031-4546-65 Age 0 5 10 15 20 25 30 35 WER (%) WER by Age (Fair-Speech) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Figure 12: WER by age group on Fair-Speech. Note the ageĂarchitecture interaction: the 46-65 group flips from worst (Whisper) to best (Qwen3). 15 LowMediumAffluent Socioeconomic 0 5 10 15 20 25 30 35 WER (%) WER by Socioeconomic Status (Fair-Speech) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Figure 13: WER by socioeconomic status on Fair-Speech. Medium-SES speakers have the highest WER across most models, with Affluent consistently lowest. EnglishSpanishMandarinHindiOther First Language (L1) 0 5 10 15 20 25 30 WER (%) WER by First Language (Fair-Speech) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Figure 14: WER by first-language group on Fair-Speech. L1-English speakers have the highest WER across most models (an ethnicity confound). Whisper models show extreme L1 sensitivity (MMR up to 15.05 across individual L1s), while Gen 3 models are substantially more robust (MMR†5.38). 16 Table 11: WER (%) by individual first language (21 L1s with n â„ 50) across ASR models on Fair-Speech. Model Arabic Cantonese Dutch English Filipino French German Hindi Indonesian Korean Malay Mandarin Marathi Nepali Other Portuguese Russian Spanish Tagalog Urdu Vietnamese MMR Gap (%) wav2vec2-large 29.6 33.4 40.8 33.1 23.3 20.7 29.8 32.6 26.0 30.2 25.5 33.3 32.7 27.1 24.4 29.6 39.0 28.9 23.7 21.6 27.7 1.97 96.7 whisper-small 4.7 5.6 4.6 12.2 1.0 14.3 10.4 9.4 4.9 2.7 2.9 5.9 4.2 2.5 9.1 4.9 5.7 11.3 3.3 5.6 4.7 15.05 1404.9 whisper-medium 4.0 4.0 3.5 9.2 2.4 8.9 8.1 4.0 5.9 2.0 3.8 4.5 1.0 1.7 6.8 3.1 4.9 10.0 2.0 5.1 4.7 9.67 866.5 whisper-large-v3 3.3 3.2 2.3 8.2 1.0 7.1 3.4 2.6 4.9 1.5 2.8 3.8 2.1 1.7 6.3 3.2 3.6 9.2 2.2 4.4 3.9 9.72 871.9 qwen3-asr-0.6b 3.8 3.2 4.2 6.4 1.0 4.3 2.7 5.8 5.2 2.6 1.1 5.4 2.9 1.9 3.9 3.1 6.1 3.7 2.2 3.4 3.4 6.74 574.3 qwen3-asr-1.7b 3.3 2.6 3.5 5.1 1.0 3.1 2.3 3.7 4.9 1.5 1.4 4.6 1.8 1.9 3.3 3.5 3.7 3.1 1.9 2.3 3.8 5.38 437.9 canary-qwen-2.5b 4.4 3.7 3.9 7.3 2.4 3.3 3.4 6.1 4.4 1.7 2.2 4.7 2.6 1.9 4.2 3.9 6.1 3.9 2.0 2.4 5.6 4.34 334.4 granite-speech-3.3-2b 3.6 5.6 4.9 9.8 2.1 4.0 5.5 6.0 5.5 3.0 3.0 6.1 5.5 2.2 9.5 4.0 22.4 5.7 2.2 3.5 5.0 10.48 948.2 granite-speech-3.3-8b 3.2 5.6 3.8 8.7 2.9 4.2 5.6 5.4 4.7 3.0 3.2 5.6 5.7 2.5 8.5 3.5 8.6 4.9 3.0 3.8 4.8 3.51 250.6 17 Table 12: Insertion rate (%) by accent across ASR models on Common Voice 24. Whisper- large-v3âs Indian-accent insertion rate (9.62%) is 6.3Ăhigher than Whisper-medium (1.53%) and dramatically exceeds all Gen 3 models (â€3.07%). ModelUSEnglandCanadaAustraliaIndianAfrican wav2vec2-large1.881.772.411.511.791.53 whisper-small1.932.451.051.003.224.59 whisper-medium1.301.160.840.801.531.15 whisper-large-v30.960.790.520.909.623.63 qwen3-asr-0.6b0.890.820.520.601.051.15 qwen3-asr-1.7b0.570.451.780.600.680.96 canary-qwen-2.5b0.720.581.880.400.720.77 granite-speech-3.3-2b0.730.870.310.200.820.78 granite-speech-3.3-8b1.740.900.630.503.071.53 G Insertion rates and hallucination categories Table 12 reports insertion rates by accent on Common Voice 24. Whisper-large-v3âs Indian- accent insertion rate (9.62%) is the most extreme value in the table, 6.3Ăhigher than Whisper- medium (1.53%) and far above any Gen 3 model (â€3.07%). This effect is accent-selective: Whisper-large-v3âs insertion rates on US (0.96%), Australian (0.90%), and Canadian (0.52%) accents are comparable to other models. Table 13 decomposes insertions on Indian-accented speech by hallucination type. Whisper- large-v3 is the only model with substantial repetition loops (43.0%), driven by autoregressive cycling on out-of-distribution acoustic input. Granite-8B shows a secondary repetition mode (70.3% of its 158 insertions), likely amplified by Q-former compression. All other modelsâ insertions are predominantly content-classified (phonetic mismatches scored as insertions by the alignment algorithm) rather than true hallucinations. Table 13: Hallucination category breakdown on Indian-accented speech (Common Voice 24, n=511 utterances, 5,154 reference words). Insertions classified as repetition (autoregressive loops), syntactic (function-word completions), or content (semantic fabrications / phonetic mismatches). CountPercentage ModelGenIns. TotalRep.Syn.ContentRep.%Syn.%Content% Wav2Vec2-large192032600.034.865.2 Whisper-small21663621011.837.360.8 Whisper-medium279030490.038.062.0 Whisper-large-v324952132364643.047.79.3 Qwen3-ASR-0.6B354017370.031.568.5 Qwen3-ASR-1.7B335011240.031.468.6 Canary-Qwen-2.5B337013240.035.164.9 Granite-Speech-2B342019230.045.254.8 Granite-Speech-8B3158111143370.38.920.9 Table 14 and Figure 15 report insertion rates by ethnicity on Fair-Speech. Notably, Granite- 2B and Granite-8B show higher insertion rates for White speakers (3.42%, 3.92%) than for Black/A speakers (2.40%, 2.63%), the opposite of all Whisper and other Gen 3 models (see §3.2). Wav2Vec2 also shows higher White than Black/A insertion rates (7.30% vs. 4.53%), but this reflects its uniformly high error rates rather than a decoder-specific pattern. Figure 16 shows the hallucination category distribution on Fair-Speech, following the same methodology as the Common Voice distribution (Figure 2 in the main text). H Inference configuration Table 15 details the inference configuration for each model. All models use greedy de- coding (no beam search, no sampling) to ensure deterministic and reproducible results. 18 Table 14: Insertion Rate (%) by Ethnicity ModelWhiteBlack/AAHispanicAsianNative AmericanPacific IslanderMiddle Eastern wav2vec2-large7.304.538.708.416.936.146.51 whisper-small0.783.371.130.852.601.191.43 whisper-medium0.711.750.800.660.970.721.03 whisper-large-v30.641.250.990.631.710.761.00 qwen3-asr-0.6b0.711.411.160.951.020.690.93 qwen3-asr-1.7b0.651.050.620.630.830.630.76 canary-qwen-2.5b0.792.001.150.831.540.881.10 granite-speech-3.3-2b3.422.401.341.681.620.880.87 granite-speech-3.3-8b3.922.632.122.851.731.071.04 White Black/A Hispanic Asian Native American Pacific Islander Middle Eastern Ethnicity 0 2 4 6 8 Insertion Rate (%) Insertion Rate by Ethnicity (FS) wav2vec2-large whisper-small whisper-medium whisper-large-v3 qwen3-asr-0.6b qwen3-asr-1.7b canary-qwen-2.5b granite-speech-3.3-2b granite-speech-3.3-8b Figure 15: Insertion rate (%) by ethnicity on Fair-Speech. Granite models show a reversed pattern: higher insertion rates for White than Black/A speakers. Text normalization is applied identically to both references and hypotheses using Whis- perâsEnglishTextNormalizer, which handles case folding, punctuation removal, number normalization (âsixââ â6â), and contraction expansion (âsheâlââ âshe willâ). Table 15: Inference configuration for all models. All use greedy decoding with identical text normalization. ModelHuggingFace IDDecodingMax TokensPrompt / Notes Wav2Vec2-large facebook/wav2vec2-large-960h-lv60-selfCTC argmaxn/aNo prompt; CTC output Whisper-small openai/whisper-smallGreedy440 language="en", task="transcribe" Whisper-medium openai/whisper-mediumGreedy440 language="en", task="transcribe" Whisper-large-v3 openai/whisper-large-v3Greedy440 language="en", task="transcribe" Qwen3-ASR-0.6B Qwen/Qwen3-ASR-0.6BGreedy256 language="English" via qwen-asr Qwen3-ASR-1.7B Qwen/Qwen3-ASR-1.7BGreedy256 language="English" via qwen-asr Canary-Qwen-2.5B nvidia/canary-qwen-2.5bGreedy256NeMo SALM API; audio locator tag prompt Granite-Speech-2B ibm-granite/granite-speech-3.3-2bGreedy256Chat template; do_sample=False, num_beams=1 Granite-Speech-8B ibm-granite/granite-speech-3.3-8bGreedy256Chat template; do_sample=False, num_beams=1 Granite prompt template. Granite-Speech models use a chat-style prompt to constrain output: System: âYou are a speech transcription system. Output ONLY the exact transcrip- tion of the audio. Do not add any commentary, explanation, or formatting.â User: â<|audio|>Transcribe this audio exactly.â A post-processing step extracts the transcription from any residual conversational wrapping. Audio preprocessing. All audio is resampled to 16 kHz mono before inference. Whis- per models use log-mel spectrogram features computed by the Whisper processor (80- dimensional for small/medium, 128-dimensional for large-v3). Wav2Vec2 operates on raw waveforms. Gen 3 models use their respective audio encoders (Qwen3: direct projection; Granite: Conformer + Q-former; Canary: FastConformer). 19 wav2vec2-large whisper-small whisper-medium whisper-large-v3 qwen3-asr-0.6bqwen3-asr-1.7b canary-qwen-2.5b granite-speech-3.3-2bgranite-speech-3.3-8b 0 20 40 60 80 100 % of Insertions Hallucination Category Distribution (FS) Syntactic Content Repetition Figure 16: Hallucination category distribution on Fair-Speech. Whisper-large-v3âs insertions are dominated by syntactic completions and repetitions; Gen 3 models have proportionally more content-classified insertions but far fewer total insertions. I Bootstrap confidence intervals Tables 16 and 17 report 95% bootstrap confidence intervals (200 resamples, utterance-level) for group-level WER on the two primary demographic axes. Non-overlapping CIs between groups indicate statistically significant differences at the 95% level. Table 16: Bootstrap 95% CIs for WER by ethnicity on Fair-Speech. Black/A CIs do not overlap with White CIs for 8 of 9 models. ModelBlack/A [95% CI]White [95% CI]Asian [95% CI]Hispanic [95% CI]Overlap? Wav2Vec2-large41.59 [40.87, 42.26]26.25 [25.47, 27.18]28.74 [27.86, 29.58]30.17 [29.13, 31.28]No Whisper-small18.57 [17.38, 20.04]10.48 [9.14, 11.63]4.59 [4.21, 5.13]7.60 [6.25, 9.09]No Whisper-medium12.42 [11.83, 12.92]9.85 [8.58, 11.13]3.78 [3.34, 4.53]6.24 [4.96, 7.49]No Whisper-large-v310.25 [9.61, 10.81]9.42 [8.21, 10.64]3.27 [2.94, 3.66]5.56 [4.54, 6.52]Yes Qwen3-ASR-0.6B11.01 [10.69, 11.33]3.00 [2.78, 3.29]3.46 [3.19, 3.78]3.84 [3.56, 4.20]No Qwen3-ASR-1.7B8.45 [8.13, 8.70]2.79 [2.57, 3.05]2.81 [2.59, 3.05]2.93 [2.70, 3.23]No Canary-Qwen-2.5B11.65 [11.20, 12.04]4.90 [4.39, 5.46]3.36 [3.11, 3.62]3.81 [3.46, 4.28]No Granite-Speech-2B13.41 [12.74, 14.30]9.96 [8.43, 11.49]5.29 [4.94, 5.80]5.62 [5.08, 6.10]No Granite-Speech-8B12.09 [11.35, 12.87]7.57 [6.18, 8.93]6.06 [5.05, 7.42]5.46 [4.96, 5.96]No Table 17: Bootstrap CIs for accent on CV 24. Whisper-large-v3âs wide Indian-accent CI [9.46, 43.53] reflects variance from hallucination episodes. ModelUS [95% CI]Indian [95% CI]African [95% CI]England [95% CI] Wav2Vec2-large16.38 [15.03, 17.51]25.17 [23.25, 27.40]23.33 [16.92, 30.79]16.45 [14.85, 18.31] Whisper-small11.48 [10.44, 12.78]17.58 [15.86, 19.54]20.27 [12.16, 29.07]13.08 [10.90, 15.11] Whisper-medium8.79 [7.95, 9.89]13.19 [11.75, 14.59]11.47 [6.89, 16.21]9.44 [7.94, 11.31] Whisper-large-v37.45 [6.69, 8.54]18.99 [9.46, 43.53]13.00 [6.51, 21.81]7.67 [6.45, 9.23] Qwen3-ASR-0.6B7.53 [6.75, 8.47]11.08 [9.69, 12.50]9.56 [5.65, 14.27]8.07 [6.84, 9.52] Qwen3-ASR-1.7B5.76 [5.15, 6.56]7.78 [6.69, 9.07]8.03 [4.83, 12.39]6.72 [5.57, 8.02] Canary-Qwen-2.5B5.84 [5.09, 6.57]7.17 [6.15, 8.24]8.03 [4.71, 12.48]6.27 [5.20, 7.73] Granite-Speech-2B8.35 [7.24, 9.41]10.36 [8.98, 11.89]9.98 [5.29, 16.22]9.04 [7.32, 11.10] Granite-Speech-8B10.63 [8.43, 13.56]12.87 [9.15, 18.02]16.06 [8.65, 23.93]9.44 [7.48, 11.64] J WER degradation curves Figure 17 shows overall WER degradation across 12 perturbation conditions for all nine models on Common Voice 24, following the Fair-Speech results shown in the main text (Figure 4). Masking is consistently the harshest perturbation; reverberation is the mildest for all models with language priors. Gen 3 models (especially Qwen3) degrade most gracefully. 20 Clean20dB10dB0dB 10 20 30 40 50 WER (%) Additive Noise Clean0.3s0.6s1.0s 10 15 20 25 Reverberation Clean25%50%75% 10 15 20 25 30 Silence Injection Clean10%20%30% 10 20 30 40 50 60 Chunk Masking WER Degradation Under Perturbation (Common Voice) W2V2-LWh-SWh-MWh-LQ3-0.6BQ3-1.7BCan-2.5BGr-2BGr-8B Figure 17: WER degradation curves on Common Voice 24. Patterns mirror Fair-Speech: Gen 3 models degrade most gracefully, masking is harshest, reverb is mildest. K Additional perturbation results Hallucination types under perturbation. Figures 18a-18c show additional hallucination classifications. The main text (Figure 6) details the distribution under masking on Fair- Speech. Under masking on Common Voice, we observe similar dynamics. W2V2-LWh-SWh-MWh-LQ3-0.6BQ3-1.7BCan-2.5BGr-2BGr-8B 0 20 40 60 80 100 Insertion Type Distribution (%) n=1,410n=25,798n=10,583n=7,888n=1,280n=993n=2,698n=3,866n=4,019 Hallucination Types Under 30% Masking (Common Voice) Repetition Syntactic (function words) Content (semantic) (a) Masking on Common Voice 24. W2V2-LWh-SWh-MWh-LQ3-0.6BQ3-1.7BCan-2.5BGr-2BGr-8B 0 20 40 60 80 100 Insertion Type Distribution (%) n=21,334n=19,289n=12,408n=12,098n=4,960n=3,916n=7,671n=10,108n=14,580 Hallucination Types Under 75% Silence (Fair-Speech) Repetition Syntactic (function words) Content (semantic) (b) Silence on Fair-Speech. W2V2-LWh-SWh-MWh-LQ3-0.6BQ3-1.7BCan-2.5BGr-2BGr-8B 0 20 40 60 80 100 Insertion Type Distribution (%) n=6,975n=15,086n=11,414n=9,392n=2,410n=1,913n=3,585n=3,311n=3,899 Hallucination Types Under 75% Silence (Common Voice) Repetition Syntactic (function words) Content (semantic) (c) Silence on Common Voice 24. Figure 18: Hallucination type distributions under masking and silence injection. The hallucination distributions in Figures 18b and 18c confirm that signal dropout, whether via masking or silence, triggers fundamentally different failure modes across generations. While Whisper-small remains dominated by repetition loops (>50%), Gen 3 explicit-LLM decoders shift their hallucination profile depending on the perturbation: under masking (Fig- ure 18a) they predominantly produce benign syntactic completions, but under severe silence injection they produce primarily content substitutions. Crucially, explicitly-pretrained LLM priors prevent the catastrophic repetition loops seen in implicit-LM systems, maintaining higher syntactic integrity even when acoustic information is entirely removed. Pareto frontier under perturbation. We provide the accuracy-fairness Pareto frontier for Common Voice accents in Figure 19. This mirrors the ethnicity findings from the main text (Figure 7): as models are stressed by masking (triangles), they move toward a âdegenerate frontier â of low accuracy and low disparity. This visualization is essential for demonstrating that the âdegradation as equalizer â effect is a universal property of autoregressive decoders rather than a dataset-specific artifact of Fair-Speech. Amplification by perturbation type. Finally, Figures 20a and 20b provide a high-level summary of the fairness gap amplification ratio (α) across all 216 inference runs, grouped 21 1015202530354045 WER (%) 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 MMR (Accent) W2V2-L Wh-S Wh-M Wh-L Q3-0.6B Q3-1.7B Can-2.5B Gr-2B Gr-8B Accuracy vs. Fairness Under Degradation (Common Voice) Condition Clean SNR 10dB Mask 20% Figure 19: Accuracy vs. accent fairness under perturbation on Common Voice 24. by perturbation type. These plots highlight a key architectural takeaway: while Chunk Masking universally compresses gaps (α<1), Silence Injection is uniquely hazardous for Whisper-smallâs accent fairness, and Additive Noise disproportionately stresses the Qwen3-1.7B accent frontier. This suggests that ârobustnessâ is not a monolithic trait but is highly sensitive to the specific type of signal corruption. W2V2-LWh-SWh-MWh-LQ3-0.6BQ3-1.7BCan-2.5BGr-2BGr-8B 0.0 0.2 0.4 0.6 0.8 1.0 Mean Fairness Amplification (MMR_pert / MMR_clean) Mean Fairness Amplification by Perturbation Type: Ethnicity (Fair-Speech) Additive NoiseReverberationSilence InjectionChunk Masking (a) Ethnicity on Fair-Speech. W2V2-LWh-SWh-MWh-LQ3-0.6BQ3-1.7BCan-2.5BGr-2BGr-8B 0.0 0.5 1.0 1.5 2.0 Mean Fairness Amplification (MMR_pert / MMR_clean) Mean Fairness Amplification by Perturbation Type: Accent (Common Voice) Additive NoiseReverberationSilence InjectionChunk Masking (b) Accent on Common Voice 24. Figure 20: Mean fairness gap amplification grouped by perturbation type. Accent WER under perturbation. Table 18 provides per-accent WER and insertion rates for the two key amplification cases discussed in §4.2. Under SNR = 10 dB, Qwen3-1.7Bâs Indian-accent WER rises disproportionately while Canadian-accent WER barely changes, drivingα= 1.63. Under 75% silence, Whisper-large-v3âs worst-affected group shifts from Indian to England: Indian-accent WER and insertion rate both decrease, while England-accent insertion rate rises from 0.8% to 22.1%. Table 18: Per-accent WER (%) and insertion rate (%) on Common Voice 24 under key perturbation conditions. Top: Qwen3-1.7B under SNR = 10 dB noise. Bottom: Whisper- large-v3 under 75% silence injection. These conditions produce the strongest fairness gap amplification (α = 1.63 and 1.38, respectively; §4.2). AfricanAustraliaCanadaEnglandIndianUSMMRα (n=51)(n=98)(n=99)(n=381)(n=511)(n=1193) Qwen3-1.7B (clean accent MMR = 1.82) WERClean8.04.44.66.77.85.81.82n/a WERSNR 10 dB11.54.43.98.011.37.42.961.63 Whisper-large-v3 (clean accent MMR = 5.38) WERClean13.05.23.67.719.27.55.38n/a WERSilence 75%13.67.14.130.414.511.37.451.38 Ins. rateClean3.60.90.50.89.61.0 Ins. rateSilence 75%2.91.20.622.12.63.6 L LLM Usage Disclosure The core ideas, experimental design, data selection and analysis in this work were solely conducted by the authors. AI assistance was used to facilitate code implementation, fix formatting, and refine the prose for clarity and flow. 22