Research·arXiv cs.CL·11h ago

Model pruning widens demographic gaps in speech recognition systems

Pruning speech models for deployment efficiency masks fairness costs that aggregate metrics fail to capture. Researchers found that audio encoder compression widens performance gaps across demographic groups in speech-to-text systems, with disparities hidden in larger models until examined closely. While LoRA fine-tuning recovers overall accuracy, it amplifies existing inequities by benefiting already-advantaged groups more. This work surfaces a critical tension in model optimization: efficiency gains achieved through standard compression pipelines can systematically degrade service quality for underrepresented populations, forcing practitioners to choose between deployment cost and equitable performance.

Modelwire context

Explainer

The paper's core finding isn't that pruning hurts minorities (known risk) but that standard evaluation metrics actively hide this degradation. Aggregate accuracy can improve or stay flat while demographic performance gaps widen, making fairness costs invisible until explicitly measured.

This connects to a broader pattern in recent work on model compression and quantization. Papers like LeapQuant and WUSH-KV have focused on maintaining reconstruction fidelity under aggressive compression, but they measure fidelity through aggregate loss or BLEU scores. This speech pruning work exposes that fidelity metrics don't capture distributional harm. Similarly, the LongHarness Bench paper from late September highlighted how standard benchmarks fail to surface real tradeoffs in production systems. Here, the tradeoff is between deployment cost and equitable access, and it's invisible to the metrics teams typically rely on.

If Fair-Speech or similar fairness-auditing frameworks are adopted into standard model evaluation pipelines at major speech vendors (Google, Meta, Apple) within the next 12 months, that signals the research has moved from academic concern to operational requirement. If pruned models continue shipping without demographic breakdowns in performance reports, the finding remains a cautionary tale rather than a practice shift.

Coverage we drew on

LeapQuant: Efficient Linear Attention with Accurate Recurrent State Quantization · arXiv cs.LG

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsSLAM-ASR · Fair-Speech · Common Voice · LoRA

Read full story at arXiv cs.CL →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Pruning for Efficiency, Paying in Fairness: Demographic Disparities in Pruned Speech-LLMs”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.