Demographic Bias in ASR

This field studies how ASR systems perform across accents, dialects, gender, age, ethnicity, and intersecting speaker groups, and how errors affect people who rely on transcription. Researchers use demographic benchmarks and in-the-wild speech, subgroup and intersectional fairness metrics, and representation probes to identify disparities that overall word error rate can hide. Mitigation approaches include balanced data, adversarial training, fairness-aware fine-tuning, and changes to model decoding or compression. Gaps persist across newer systems, while removing demographic information or improving average accuracy does not reliably improve fairness; results can also depend on the metric, dataset, and training seed.