"This Wasn't Made for Me": Recentering User Experience and Emotional Impact in the Evaluation of ASR Bias

2026-04-24 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A 2026 study investigated the human experience of bias in Automatic Speech Recognition (ASR) systems, moving beyond error rates to examine emotional and psychological impacts. Researchers conducted user experience studies across four U.S. locations (Atlanta, Gulf Coast, Miami Beach, Tucson) representing distinct English dialect communities, surveying 215 participants. Findings indicate that over 53% of participants reported technologies failing to consider their cultural backgrounds, requiring constant adjustment for basic functionality. Despite these challenges, over 80% maintained high expectations for ASR performance, and 75.5% expressed willingness to contribute to model improvement. Qualitative analysis revealed participants experience frustration, annoyance, and feelings of inadequacy, often internalizing system failures as personal shortcomings despite recognizing that systems were not designed for them. Users perform extensive "invisible labor," including code-switching, hyper-articulation, and emotional management, to make ASR functional, highlighting a significant gap between technical fairness metrics and lived user harm.

Key takeaway

For AI Product Managers and Research Scientists developing ASR, your current fairness assessments based solely on accuracy metrics are incomplete. You must integrate user experience and emotional impact evaluations to understand the full scope of harm, including the "invisible labor" and psychological toll on users from underrepresented dialect communities. Prioritize participatory design frameworks that equitably compensate and empower these communities, rather than merely extracting data, to build truly inclusive voice technologies.

Key insights

ASR bias creates significant emotional and cognitive burdens for users of non-dominant dialects, extending beyond mere accuracy issues.

Principles

• Accuracy metrics alone fail to capture the full scope of algorithmic harm.
• Users often internalize system failures as personal inadequacy, despite critical awareness.
• Technological exclusion imposes invisible labor on marginalized communities.

Method

A mixed-methods approach combined structured questionnaires and open-ended narrative responses from 215 participants across four U.S. dialect communities to analyze user experiences and emotional impacts of ASR bias.

In practice

• Evaluate ASR systems using user experience and emotional impact metrics.
• Compensate users for data contributions to improve biased systems.
• Design ASR with participatory frameworks centering affected communities.

Topics

• ASR Bias
• User Experience
• Emotional Impact
• Linguistic Discrimination
• Invisible Labor

Best for: Research Scientist, AI Product Manager, AI Scientist, NLP Engineer, AI Ethicist

Related on AIssential

• Transcribing speech is never neutral. It shapes power and bias — Language technologies trained on prestige dialects often misrepresent non-standard and Indigenous communication, leading to serious consequences.
• Imperfectly Cooperative Human-AI Interactions: Comparing the Impacts of Human and AI Attributes in Simulated and User Studies — AI transparency significantly impacts human trust and satisfaction in imperfectly cooperative interactions, often more than human personality.
• People Hate AI Even More Than They Hate ICE, Poll Finds — Public sentiment towards AI is predominantly negative, with many users disliking its societal impacts despite personal use.
• Algorithms don’t care: how AI worsens the double burden for Indonesia’s female gig workers — AI platforms can perpetuate colonial exploitation by embedding extractive logics that disadvantage women in the Global South's gig economy.
• The Voice Behind the Words: Quantifying Intersectional Bias in SpeechLLMs — SpeechLLMs exhibit implicit intersectional bias, providing less helpful responses to specific accent-gender combinations, requiring human evaluation for full detection.
• BAFIS: Dataset + Framework to assess occupational Bias and Human Preference in modern Text-to-image Models — Text-to-image models exhibit systematic occupational biases, requiring human preference feedback for fair evaluation and development.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.