Toward Fair Speech Technologies: A Comprehensive Survey of Bias and Fairness in Speech AI

Yi-Cheng Lin and colleagues from multiple institutions have released a landmark survey on bias and fairness in speech AI, covering over 400 studies across generation and perception tasks. The paper addresses a critical gap: existing fairness surveys either take a general machine learning view that overlooks speech-specific properties, or focus on single tasks, missing shared failure patterns. The authors propose a unified framework that links formal fairness definitions to evaluation, diagnosis, and mitigation. They formalize seven fairness definitions adapted to speech—including demographic parity, equal opportunity, and individual fairness—and organize the field's conceptual evolution through three paradigms: Robustness (ensuring consistent performance across groups), Representation (fair treatment in model outputs), and Governance (oversight and accountability).

To help practitioners select appropriate metrics, the survey grounds evaluation measures in the mathematical cores of these definitions and offers a decision tree. It then diagnoses bias sources along the entire speech processing pipeline, surfacing speech-specific mechanisms such as channel bias (e.g., microphone quality as a demographic proxy) and annotation subjectivity in emotion labels. Mitigation strategies are systematized across four intervention stages: data, preprocessing, model training, and post-processing, each mapped to diagnosed sources. The survey identifies open challenges—like fairness in emerging speech-language models—and proposes future research directions. This work is a must-read for any team building or deploying speech technologies in high-stakes settings such as healthcare, hiring, or law enforcement.