Bias Mitigation in Audio-Enabled LLMs through Adversarial Training and Distillation Methods

1 Introduction

• Vocal cues can reveal attributes such as gender and lead AuLLMs to produce stereotyped outputs for identical recording content.The paper frames fairness as important as AuLLM adoption grows, including among people with limited digital literacy.Source: §1 · p. 1
• The authors propose two successive methods for AuLLM bias mitigation, demonstrating them on gender while presenting a framework applicable to attributes such as age and group-specific accents.Figure 1 outlines the adversarial-only method and the full adversarial/distillation pipeline.Source: §1 · p. 1, Figure 1 · p. 1
• Adversarial training removes gender signals from the audio encoder to create a gender-unaware model, while distillation aims to retain gender perception with comparable outputs across masculine and feminine voices when gender is irrelevant.The gender-unaware model retains audio understanding comparable to the original, according to the authors.Source: §1 · p. 2
• The contributions include examining side effects of gender removal, presenting a bias-mitigation approach that preserves gender perception, and releasing code and trained models.The authors also describe the topic of bias mitigation in AuLLMs as understudied.Source: §1 · p. 2

2 Related Work

• Text-only LLMs have been observed to reproduce and amplify stereotyped narratives, and current models still exhibit undesirable behavior across several attributes.The cited work covers gender, race, age, disability, and sexual orientation, among others.Source: §2 · p. 2
• AuLLMs remain less studied than text-only LLMs; prior evaluation found that they infer attributes such as gender from paralinguistic cues and produce stereotyped outputs even without explicit attribute prompts.The behavior was reported across open- and closed-weight models, and in option-choice and open-generation settings.Source: §2 · p. 2
• Text-model mitigation methods such as RLHF, PEFT, and adversarial training are rarely directly applicable to AuLLMs because of architectural differences, scarce multimodal data, and differing bias manifestations.The cited approaches include RLHF, PEFT on augmented data, and adversarial training.Source: §2 · p. 2
• Pitch modification can shift generated terms toward gender stereotypes, but may degrade pitch, sound-identification, or gender-detection performance and may not eliminate gender perception.The authors report confirming these concerns and choose to adapt adversarial learning and PEFT instead.Source: §2 · p. 2

3 Datasets, Evaluation Metrics, Models

• Corpus annotations provide speaker-gender labels, and the study uses multiple datasets to train or evaluate gender detection, gender bias, and general audio capabilities.The models are Qwen2-Audio-7B-Instruct, Qwen3-Omni-30B-A3B, and Nemotron-3-Nano-Omni-30B-A3B, with chain-of-thought disabled for comparability.Source: §3 · p. 2, §3.6 · p. 4

3.1 Gender Detection

• Gender detection uses 500 British Dialects and 470 Speech Archive recordings; models choose M, F, or X, and evaluation reports absolute accuracy, refusal rate, and accuracy on non-refused samples.The corpora contain short English recordings of speakers varying in age and accent and are balanced by gender.Source: §3.1 · p. 2

3.2 Gender Bias Evaluation

• Gender-bias evaluation covers restricted-option choices and open generation; option-choice LOR summarizes stereotypical word associations, while Fightin’ Words identifies statistically significant adjective associations in generated profiles.Open-generation prompts compare profiles and gender-polarized roles for paired recordings with the same linguistic content; Figure 2 displays Nemotron-3-Nano-Omni adjectives sorted by LOR.Source: §3.2 · pp. 2–3, Figure 2 · p. 3

3.3 Misgendering Evaluation

• MISGEND measures gendered pronouns in model descriptions of voices after recordings with gendered pronouns in their transcriptions are excluded.Responses are classified as M, F, X for neutral they, or _ for no pronouns or a tie.Source: §3.3 · p. 3

3.4 Implicit Gender Stereotype Signals

• IMPGENDER pairs VCTK recordings with natural prompts where stereotypical answers are possible but not stated, providing a separate evaluation of implicit gender-stereotype signals.The dataset contains 8,160 samples, with 240 random recordings per prompt.Source: §3.4 · p. 3

3.5 General Audio Tasks

• For general audio tasks, gender-related prompts are filtered from a VCTK-only SIFT50M subset to limit interference with unlearning, and MMAU evaluates general audio understanding.The authors note that gender information may remain implicitly, for example in pronouns; the ungendered MMAU set contains 908 samples.Source: §3.5 · pp. 3–4

3.6 Models

• The three models use Whisper, AuT, and Parakeet audio encoders, respectively; all predict gender well but also produce gender-stereotypical outputs.Qwen2-Audio, the older and smaller model, exhibits more bias and considerably worse audio understanding.Source: §3.6 · p. 4

4 Adversarial Training

• The gender-unaware model removes gender representation from the voice encoder using adversarial training and a gradient reversal layer that reverses the gradient by a negative constant during backpropagation.The method aims to remove the gender signal from the encoder output.Source: §4 · p. 4

4.1 Experimental Setup

• Training updates only the audio encoder, mean-pools its final representations, and sends them through a gradient reversal layer to a binary gender head for cross-entropy training.A language-model cross-entropy loss is combined with the gender loss using hyperparameter γ to help prevent catastrophic forgetting.Source: §4.1 · p. 4
• The authors trained on 4,000 gender-balanced samples for one epoch at a constant learning rate, checkpointing the encoder every 125 optimization steps; they selected checkpoints below 60% gender-detection accuracy and chose the one with highest MMAU benchmark accuracy.Source: §4.1 · p. 4

4.2 Training and Evaluation

• All three models reached near-random gender detection; general audio understanding degraded by <4% for Qwen3-Omni and Nemotron-3-Omni, while Qwen2-Audio improved and Qwen3-Omni’s MMAU results remained comparable to its original model.The table presents selected hyperparameter configurations; Qwen3-Omni was hardest to train, with only two checkpoints below 60% gender detection accuracy.Source: §4.2 · pp. 4–5, Table 1 · p. 5
• Near-random gender-detection accuracy did not always indicate genuine uncertainty: Qwen3-Omni usually predicted female, while other models sometimes refused and retained up to 80.8% accuracy on non-refused samples.The authors identify qualitative analysis as important for uncovering such shortcuts.Source: §4.2 · p. 5, Table 2 · p. 5
• All gender-unaware models reduced overall bias LOR, but their outputs became less balanced, with Qwen2-Audio and Qwen3-Omni favoring stereotypically masculine options and Nemotron-3 favoring feminine options.On MISGEND, gender unlearning increased misgendering or pronoun avoidance; Nemotron-3-Nano still used the correct pronoun in 72% of cases.Source: §4.2 · p. 5
• Gender-unaware Qwen3-Omni and Nemotron-3-Nano produced shorter MISGEND responses, while Qwen2-Audio responses grew longer; the authors link brevity possibly to concise SIFT-50M targets averaging 75.2 characters.The reported length comparisons are against each model’s original variant.Source: §4.2 · p. 5

5 Distillation Stage

• Distillation combines logits from the original model and its gender-unaware variant to train a model that perceives gender but behaves gender-agnostically when gender is irrelevant.The target model is fine-tuned using low-rank adaptation on decoder attention weights.Source: §5 · pp. 5–6, §5.1 · p. 6

5.1 Experimental Setup

• The setup mixes up to 4,000 samples from each model-generated dataset, then trains for one epoch using length-normalized KL-divergence loss.The two datasets use original-model answers to SIFT-50M prompts and GUM answers to IMPGENDER prompts.Source: §5.1 · p. 6

5.2 Training and Evaluation

• The authors grid-searched eight hyperparameter combinations with ten checkpoints each and selected checkpoints meeting gender-detection and MMAU-accuracy thresholds.Table 4 presents the top three qualifying checkpoints per model, sorted by LOR.Source: §5.2 · pp. 6–7, Table 4 · p. 7
• Qwen2-Audio’s gender-bias LOR dropped considerably, while Qwen3-Omni and Nemotron-3-Omni showed only slight improvement; most checkpoints for the latter two scored worse than their originals.The reported proportions of worse-scoring checkpoints were 86.3% for Qwen3-Omni and 83.8% for Nemotron-3-Omni.Source: §5.2 · p. 7, Table 4 · p. 7
• Distilled models’ gender detection resembled their originals, except Qwen2-Audio refused to classify more masculine voices than before.Checkpoint selection required gender-detection accuracy of at least 70% and MMAU accuracy no worse than 97% of original performance.Source: §5.2 · p. 7
• Bias patterns remained model-specific: Qwen2-Audio leaned masculine, Qwen3-Omni resembled its original with slightly weaker bias, and Nemotron-3-Nano-Omni leaned feminine less strongly than its GUM.Open-generation analysis found less stereotypically charged word associations for Qwen2-Audio and a smaller positive improvement for Qwen3-Omni; Nemotron’s fewer associated words may reflect shorter answers rather than balanced outputs.Source: §5.2 · p. 7
• Across all three models, distilled-model misgendering rates were comparable to those of their original variants.The authors report this as consistent with gender-detection results.Source: Figure 4 · p. 6, §5.2 · p. 7
• Response lengths shifted differently: Qwen2-Audio’s final average fell between its original and GUM, Qwen3-Omni recovered its original level, and Nemotron-3-Nano-Omni became shorter than both.The authors relate these patterns partly to the custom SIFT-50M model-generated dataset; its average response lengths were 96.6, 303.0, and 85.5 characters, respectively.Source: Figure 5 · p. 6, §5.2 · p. 7

6 Discussion

• Near-random gender accuracy can hide shortcuts: Qwen3-Omni chose the ‘F’ label 95.5% of the time, while the other models often refused but frequently answered correctly otherwise.The authors argue that qualitative analysis is needed because absolute accuracy alone can mislead.Source: §6 · pp. 7–8
• Most distilled Qwen3-Omni and Nemotron-3-Nano-Omni checkpoints were more biased than their originals, which the authors attribute to SIFT-50M’s factual-skill focus and samples rewarding stereotypical outputs.A non-trivial subset of checkpoints showed slight improvement.Source: §6 · p. 8
• Even generated answers from richer-text models can remain short and on-point because the prompts elicit ‘benchmarking mode’; the authors also note error accumulation and stereotype-enforcing questions.This constrains the proposed use of model-generated answers as a more suitable distillation dataset.Source: §6 · p. 8
• Selective and parameter-efficient training keeps the methods’ dataset requirements relatively small; the authors suggest a carefully curated dataset of about 15k samples could be feasible.Source: §6 · p. 8

• Adversarial training largely removed gender perception, though some models retained it or confidently claimed to perceive gender they no longer perceived.The resulting models target settings where gender inference is undesired.Source: §7 · p. 8
• Parameter-efficient logit distillation combined original and gender-unaware outputs, selecting checkpoints with comparable gender perception and general audio understanding and the least bias.Source: §7 · p. 8
• Improvement was marginal for two of three models, and most distilled checkpoints were more biased than the original baseline.Source: §7 · p. 8

• Available prompts are short and factual, may reward assumptions over acknowledging limited knowledge, and include samples that encourage biased behaviour; suitable alternatives are scarce.Source: Limitations · p. 9
• IMPGENDER uses 34 unique prompts repeated 240 times; the authors describe hundreds of diverse prompts repeated only a few times as ideal, but resource-intensive to create.Source: Limitations · p. 9
• Evaluation presents a sample with unrelated content and a prompt, unlike real use where the instruction is often spoken in the recording.Source: Limitations · p. 9
• The study covers English alone, and the authors note that gendered languages may pose new challenges, especially for filtering gendered prompts.Source: Limitations · p. 9
• Gender labels are limited to male and female, which the authors say does not reflect the complexity of gender identities.Source: Limitations · p. 9

• The authors state that they do not anticipate negative ethical implications arising from the study.Source: Ethical Considerations · p. 9
• GitHub Copilot was used for AI-assisted coding, while Claude, ChatGPT, and Grammarly supported grammar checking and minor rephrasing.Source: Ethical Considerations · p. 9
• Qwen2-Audio and Qwen3-Omni use Apache 2.0, Nemotron-3-Nano-Omni uses the NVIDIA Open Model Agreement, and the listed corpora and datasets have separate licenses.The listed licenses include CC-BY-SA-4.0, the Edinburgh DataShare depositor agreement, CDLA-SHARING-1.0, and CC-BY-NC-4.0.Source: Ethical Considerations · p. 9

A Stereotypes in SIFT-50M

• Manual inspection found SIFT-50M prompts eliciting gender stereotypes, including feminine speakers framed as nurturing and young women as less authoritative.Source: A Stereotypes in SIFT-50M · p. 11
• An Irish accent is described as adding cultural and regional flavor that can shape perceived background and credibility.Source: A Stereotypes in SIFT-50M · p. 11

B Technical Details

• Training used Transformers, a custom loop, AdamW, and Accelerate; vLLM served inference, one epoch took about two hours, and Table 5 reports per-model GPU memory use.Source: B Technical Details · p. 12, Table 5 · p. 12

B.1 Adversarial Training Hyperparameters

• Adversarial training used batch size 8 and weight decay 5 × 10−3, while learning rate, LLM-loss weight, and gradient-reversal weight were grid-searched.The searched values were learning rates 5 × 10−6, 10−5, 5 × 10−5, and 10−4; LLM-loss weights 0.5, 1, and 2; and reversal weights 0.025, 0.05, 0.1, and 0.2.Source: §B.1 · p. 12

B.2 Distilling Hyperparameters

• Distillation used batch size 8, weight decay 5 × 10−3, LoRA dropout 0.05, and LoRA alpha twice the LoRA rank.Learning rate and LoRA rank were grid-searched, with ranks 16 and 32.Source: §B.2 · p. 12

C Prompts

• The prompt set includes a gender-detection question with M, F, or X response options, a voice-description prompt, and a list of gender-related words filtered from ungendered datasets.A separate listing is identified as the prompt for misgendering detection.Source: C Prompts · p. 12

C.1 Implicit Gender Stereotype Prompts

• Implicit-stereotype prompts ask models to infer roles, performance, leadership, and social explanations from voices, including scenarios involving gender-polarised contexts.Examples include guessing a person’s role at a workplace or institution and assessing a student’s performance or leadership.Source: §C.1 · pp. 12–13
• After distillation, Qwen2-Audio’s open-generation adjectives were described as less stereotypically charged than before training.Source: Figure 9 · p. 14, Figure 8 · p. 14
• For Qwen3-Omni, the post-distillation adjective comparison showed minor improvement in the power subplot and LOR values on the x axis.Source: Figure 11 · p. 14, Figure 10 · p. 14