Pruning for Efficiency, Paying in Fairness: Demographic Disparities in Pruned Speech-LLMs

Ganesh Pavan Kartikeya Bharadwaj Kolluri; Michael Kampouridis; Ravi Shekhar

cs.SDSep 29, 2026

Pruning for Efficiency, Paying in Fairness: Demographic Disparities in Pruned Speech-LLMs

Authors: Ganesh Pavan Kartikeya Bharadwaj Kolluri, Michael Kampouridis, Ravi Shekhar

Organizations: School of Computer Science and Electronic Engineering, University of Essex, UK

Abstract

Speech-LLMs are expensive to run, making compression important for real-world deployment. However, compressed models are usually selected using aggregate word error rate (WER), which can hide how pruning affects different demographic groups. In this work, we systematically study the effect of audio encoder pruning on SLAM-ASR for different demographic groups. Using the Fair-Speech and Common Voice datasets, we found that the pruning does not affect all demographic groups equally; the gap between best- and worst-performing groups increases in fold. These disparities appear across all three encoder scales, but only the largest model initially hides them behind aggregate WER. LoRA adaptation improves WER for every group, but benefits groups already performing well more strongly and widens for certain groups. On Common Voice English, Danish, and Dutch, accent gaps persist but do not clearly widen, showing that the fairness effects of pruning vary across datasets and must be measured directly. Our findings suggest that for pruned models, deployment decisions should include per-group WER, with the worst-performing group's error rate as an explicit criterion.

Figures & tables

Figure 1: Compression widens the racial gap, and the aggregate WER hides it. WER for Black and Asian speakers (Fair-Speech, Whisper large) under top-down pruning; the wedge between the curves is their gap, which nearly doubles (

13.5\!\rightarrow\!24.5

pp). Within the grey band, aggregate WER improves (

21.6\!\rightarrow\!21.1

%), yet Black speakers are the only group whose word error rate rises significantly (

+0.9

pp,

p{=}.009

Corpus	Split	Samples
Training (Common Voice 22)
English (EN)	Train	58,140
Dutch (NL)	Train	43,458
Danish (DA)	Train	3,592
Evaluation
Common Voice (EN)	Test	16,391

Table 1: Dataset statistics. Common Voice 22 provides the training and test splits; Fair-Speech is used for evaluation only.

Corpus	Bias axes
CV-22 (EN/DA/NL)	accent, gender, age
Fair-Speech	ethnicity, SES, gender, age, L1

Table 2: Evaluation corpora and the demographic axes each supports.

Group	$n$	L-0	L-2	L-4	L-6	L-8
Asian	3,854	$13.7$	$13.1$	$16.3$	$22.2$	$26.5$
Native Haw.	969	$13.4$	$13.7$	$17.8$	$24.7$	$31.4$
Hispanic	2,811	$20.0$	$19.2$	$22.5$	$27.5$	$33.9$
White	5,619	$20.6$	$20.1$	$22.7$	$28.4$	$31.7$
Native Am.	4,616	$21.5$	$19.0$	$22.1$	$29.0$	$33.5$
MENA	749	$23.0$	$22.3$	$26.0$	$34.2$	$37.3$

Table 3: Fair-Speech, ethnicity (Whisper large);

n

is the number of utterances per group. Black speakers have the highest WER at every depth. At L-2 the aggregate improves while Black speakers degrade, and by L-8 the Black–Asian gap has nearly doubled, from 13.5 to 24.5 pp.

Figure 2: Worst-best WER gap per demographic axis under top-down pruning (Fair-Speech, Whisper large); one line per axis, with the pair named in the legend. The ethnicity and gender gaps widen, the Socioeconomic gap holds, and the age gap narrows by leveling down.

Figure 3: Per-group WER on Fair-Speech as encoder layers are removed, per Whisper scale (shared

y

-axis; dashed = aggregate). Black speakers have the highest error rate at every scale and every depth. Only at large-v2 does the shallow prune leave the aggregate unchanged. Appendix Table 11 covers that depth.

Figure 4: Change in WER (percentage points) per subgroup as Whisper large-v2 is pruned top-down, without (a) and with (b) LoRA (Fair-Speech). Warm colors indicate degradation, cool colors indicate improvement, and the bottom row is the aggregate. Black speakers degrade most at nearly every depth in both conditions. LoRA lowers the aggregate everywhere, but it improves only the already better-performing groups: at ten removed layers it recovers $8.0$ pp for Asian speakers and only $3.3$ pp for Black speakers.

Model	L-0	L-2	L-4	L-6	L-8
small (12L)	2.03	1.70	1.51	1.49	1.13
medium (24L)	2.16	2.01	1.94	1.97	1.66
large-v2 (32L)	1.99	2.15	2.10	1.93	1.92

Table 4: Fair-Speech, disparity ratio

\rho

between Black and Asian speakers, per encoder scale. Italic values lie beyond the usable range (aggregate WER

>40\%

). Only large-v2 shows a rise at the first prune, where its aggregate WER also improves.

Group	$n$	L-0	L-2	L-4	L-6	L-8
US	1,145	$11.1$	$12.0$	$13.9$	$16.7$	$17.8$
England	358	$12.2$	$13.2$	$15.1$	$17.1$	$18.9$
India/S-Asia	508	$15.8$	$17.8$	$19.1$	$21.0$	$24.8$
Aggregate	3,008	$12.1$	$13.0$	$14.7$	$17.0$	$19.1$

Table 5: Common Voice 22 English, accent (Whisper large);

n

is the number of utterances per group. India/South Asia has the highest WER at every depth. Unlike Fair-Speech, the aggregate degrades from the first prune.

Group	$n$	L-0	L-2	L-4	L-6	L-8
Netherlands	3,817	$11.5$	$12.9$	$15.0$	$18.1$	$20.4$
Belgian	1,146	$13.7$	$15.2$	$17.8$	$20.5$	$25.2$
Aggregate	5,306	$11.8$	$13.3$	$15.5$	$18.5$	$21.4$

Table 6: Common Voice Dutch, accent (Whisper large);

n

is the number of utterances per group. Belgian speakers have higher WER than Netherlands speakers at every depth.

Group	$n$	L-0	L-2	L-4	L-6	L-8
Male	3,372	$13.5$	$15.3$	$17.7$	$18.5$	$20.6$
Female	1,163	$10.4$	$11.1$	$13.3$	$14.8$	$15.9$
Aggregate	4,535	$12.7$	$14.2$	$15.5$	$16.5$	$19.4$

Table 7: Common Voice Dutch, gender (Whisper large);

n

is the number of utterances per group. Male speakers have higher WER at every depth.

Axis	Group	$n$	L-0	L-2	L-4	L-6	L-8
Gender	Male	1,031	$36.0$	$38.3$	$40.6$	$49.0$	$51.8$
Gender	Female †	299	$33.0$	$34.1$	$35.2$	$42.9$	$48.0$
	Aggregate	1,330	$35.3$	$37.4$	$39.4$	$47.6$	$50.9$
Age	Teens †	50	$40.3$	$42.0$	$43.6$	$49.9$	$53.6$
	Twenties	601	$41.1$	$43.2$	$45.2$	$53.7$	$56.3$
	Thirties	392	$31.5$	$34.2$	$36.8$	$45.0$	$48.3$

Table 8: Common Voice Danish (Whisper large);

n

is the number of utterances per group. Groups marked

\dagger

fall below the analysability threshold and are reported for coverage only.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Language	Batch	Proj.	LoRA
English	8	2	4
Dutch	8	4	8
Danish	4	6	8

Appendix

Table 9: Batch size and epoch counts per language. Settings are identical across all three encoder scales and all pruning depths.

Group	$n$	L-0	L-2	L-4	L-6	L-8
Asian	3,854	1.00	0.95	1.19	1.62	1.93
Native Haw.	969	1.00	1.02	1.33	1.84	2.34
Hispanic	2,811	1.00	0.96	1.12	1.37	1.69
White	5,619	1.00	0.98	1.10	1.38	1.54
Native Am.	4,616	1.00	0.88	1.03	1.35	1.56
MENA	749	1.00	0.97	1.13	1.49	1.63

Appendix

Table 10: Fair-Speech, WER relative to each group’s own unpruned baseline (Whisper large);

n

is the number of utterances per group. Values below 1.00 indicate improvement. At L-2 the aggregate is at 0.98 while Black speakers are at 1.03, the highest of any group.

Model	Series	$n$	L-0	L-2	L-4	L-6	L-8
large-v2 (32L)	Asian	3,854	1.00	0.95	1.19	1.62	1.93
	Black	7,799	1.00	1.03	1.26	1.57	1.87
	Aggregate	26,417	1.00	0.98	1.17	1.48	1.74
medium (24L)	Asian	3,854	1.00	1.17	1.51	1.72	2.64
	Black	7,799	1.00	1.09	1.35	1.57	2.03
	Aggregate	26,417	1.00	1.11	1.36	1.52	2.10

Appendix

Table 11: Fair-Speech, WER relative to each model’s own unpruned baseline;

n

is the number of utterances. Italic values lie beyond the usable range (aggregate WER

>40\%

). Only large-v2 has a depth where the aggregate improves while Black speakers degrade.

Group	$n$	L-0	L-2	L-4	L-6	L-8
Asian	3,854	$12.5$	$14.7$	$18.8$	$21.4$	32.9
Native Haw.	969	$14.8$	$17.6$	$20.3$	$24.1$	36.1
White	5,619	$18.5$	$21.7$	$24.9$	$28.2$	38.9
Hispanic	2,811	$18.7$	$21.9$	$25.0$	$26.0$	39.4
Native Am.	4,616	$19.3$	$22.7$	$26.0$	$26.5$	38.5
MENA	749	$22.8$	$25.5$	$28.2$	$33.1$	42.7

Appendix

Table 12: Fair-Speech, ethnicity (Whisper medium);

n

is the number of utterances per group. Italic values lie beyond the usable range (aggregate WER

>40\%

Group	$n$	L-0	L-2	L-4	L-6	L-8
Asian	3,854	$17.9$	$28.9$	45.8	51.9	91.9
Native Haw.	969	$18.7$	$31.2$	51.8	60.9	101.8
White	5,619	$23.3$	$32.9$	51.4	56.9	94.0
Native Am.	4,616	$23.6$	$34.0$	53.5	60.5	100.8
Hispanic	2,811	$24.5$	$33.0$	51.6	57.8	98.6
MENA	749	$30.8$	$39.2$	55.9	61.8	103.4

Appendix

Table 13: Fair-Speech, ethnicity (Whisper small);

n

is the number of utterances per group. Italic values lie beyond the usable range (aggregate WER

>40\%

Axis	Group	$n$	Minutes
Ethnicity	Black	7,799	1,010
	White	5,619	702
	Native Am.	4,616	556
	Asian	3,854	462
	Hispanic	2,811	372
	Native Haw.	969	94

Appendix

Table 14: Fair-Speech coverage per demographic group: utterances and minutes of audio. Every analyzed group exceeds the analysability threshold of 200 utterances and 30 minutes.

Figure 5: WER by gender under top-down pruning (Fair-Speech, Whisper large-v2; dashed = aggregate). Male speakers trail at every depth, and the gap widens from 3.3 pp at L-0 to 11.1 pp at L-8: female speakers improve slightly at the first prune while male speakers do not, and they degrade more slowly thereafter.

Figure 6: WER by socioeconomic group under top-down pruning (Fair-Speech, Whisper large; dashed = aggregate). Affluent speakers keep an advantage at every depth. The gap does not amplify: the Low-to-Affluent ratio falls from 1.36 to 1.24 as all groups degrade.

Axis	Group	$n$	L-0	L-2	L-4	L-6	L-8
SES	Low	14,741	$22.3$	$21.5$	$25.2$	$31.9$	$37.2$
	Medium	9,809	$21.5$	$21.6$	$26.4$	$33.2$	$39.5$
	Affluent	1,867	$16.4$	$15.7$	$18.2$	$25.5$	$30.0$
Gender	Male	12,035	$23.4$	$23.7$	$28.9$	$36.6$	$43.7$
Gender	Female	14,382	$20.1$	$19.1$	$22.2$	$28.2$	$32.6$
Age	18–22	3,845	$16.2$	$15.5$	$18.8$	$25.8$	$31.0$

Appendix

Table 15: Fair-Speech, socioeconomic status, gender, and age (Whisper large);

n

is the number of utterances per group.

Figure 7: WER by age group under top-down pruning (Fair-Speech, Whisper large-v2; dashed = aggregate). The oldest group (46–65), worst at baseline, degrades slowest (

+9.9

pp by L-8), while the 31–45 group degrades fastest (

+19.8

pp) and overtakes it as the worst-performing group by L-6. The worst-best gap narrows, but by leveling down, not by helping the group that started worst.

Group	$n$	L-0	L-2	L-4	L-6	L-8	L-10
Asian	3,854	$11.1$	$11.2$	$14.2$	$18.5$	$22.2$	$29.1$
Native Haw.	969	$13.8$	$14.4$	$16.0$	$21.3$	$27.0$	$36.7$
Native Am.	4,616	$14.4$	$15.6$	$18.0$	$22.2$	$26.2$	$34.8$
Hispanic	2,811	$15.6$	$16.1$	$19.6$	$22.8$	$28.6$	$33.5$
White	5,619	$16.6$	$17.9$	$18.8$	$23.8$	$25.8$	$32.9$
MENA	749	$17.9$	$21.0$	$23.1$	$27.6$	$30.6$	$40.0$

Appendix

Table 16: Fair-Speech, ethnicity under LoRA adaptation (Whisper large);

n

is the number of utterances per group. WER is lower than in the base model (Table 3 ) for every group at every depth.

Figure 8: Common Voice Dutch: Belgian vs Netherlands accent WER under top-down pruning, per Whisper scale (usable range only; dashed = aggregate). Belgian speakers trail at every scale and depth, but the ratio stays between 1.1 and 1.3 throughout: the gap does not amplify. Each panel ends where its model leaves the usable range.

Explore similar work

Temporal Taxation Compounds Under Post-Training Compression of Whisper Models
Sep 23, 2026Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers +1Whisper-Large-V3 Character Error Rate
The Ohio State University, Columbus, OH, USA · Air Force Research Laboratory, USA
Do LLM Decoders Listen Fairly? Benchmarking How Language Model Priors Shape Bias in Speech Recognition
Apr 23, 2026Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers +1Large Language Model Decoding Algorithmic Fairness
The Ohio State University, Columbus, OH, USA · Air Force Research Laboratory, USA
When Entanglement Lower-Bounds Disparity: Auditing and Repairing Demographic Fairness in Audio Understanding Models
Sep 23, 2026Kian Shamsaie, Iman ModarressiSpeaker Disparities
People Make Things

Sep 23, 2026, cs.CLJ/K move · Enter open · S save

Temporal Taxation Compounds Under Post-Training Compression of Whisper Models

Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers, Srinivasan Parthasarathy

The Ohio State University, Columbus, OH, USA · Air Force Research Laboratory, USA

Whisper-Large-V3 Character Error Rate Algorithmic Fairness Language Model Perplexity Post-Training

Open paper AskPDF arXiv

Automatic speech recognition models are audited for demographic fairness at full precision, yet the models that ship to production have been quantized, pruned, and distilled. We ask whether post-training weight compression, which alters model weights rather than the audio signal or its feature representation, redistributes error burden across demographic groups. Across the Whisper family on Fair-Speech, Common Voice 25, and AfriSpeech-200, 50% Wanda pruning of Whisper-large-v3 sharply widens the Black/AA-vs-Asian temporal-taxation differential on Fair-Speech: the absolute word-error-rate gap between the worst- and best-served groups more than doubles; at an assumed cost of five seconds of correction effort per transcription error this is a rise from 30 to 64 seconds of correction time per minute of speech. This +111% relative increase is invariant to the assumed per-error cost, survives an audio-quality control, and is only partly mitigated by beam-search decoding, which still leaves an +86% increase. At edge model size, INT4 HQQ quantization compounds catastrophic transcript loops on West African accents by factors of five to seven. Distillation, by contrast, narrows demographic gaps in 21 of 27 evaluated settings (teacher-student pair, precision, and dataset), with the exceptions concentrated on a single model pair. We cast the temporal-taxation construct of Choi and Choi (2025) as a quantitative metric, and show that single-snapshot fairness audits on full-precision models do not capture the deployment-time burden that compression places on already-marginalized speakers.

Abstract

Explore similar work

Temporal Taxation Compounds Under Post-Training Compression of Whisper Models

Do LLM Decoders Listen Fairly? Benchmarking How Language Model Priors Shape Bias in Speech Recognition

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