Identifying and typifying demographic unfairness in phoneme-level embeddings of self-supervised speech recognition models

Technical Summary: Identifying and Typifying Demographic Unfairness in Phoneme-Level Embeddings

Problem Statement

Despite significant advancements in Automatic Speech Recognition (ASR), modern systems exhibit persistent demographic unfairness, performing significantly better for certain Speaker Groups (SGs) than others. Current research often treats ASR models as "black boxes," focusing on improving overall fairness without understanding the underlying modeling errors. This paper seeks to move beyond black-box analysis by investigating whether demographic unfairness in speech encoders stems from systematic error (embedding bias) or random error (unequal variance) at the phoneme level.

Methodology

The authors propose a framework to distinguish between two types of modeling errors:

1. Embedding Bias (Systematic Error): Phonemes are modeled around different "modes" (centers) depending on the SG. A classifier trained on one SG would fail on another because the representations are shifted.
2. Unequal Variance (Random Error): Phonemes are modeled around the same modes across SGs, but certain SGs exhibit much higher noise/dispersion in their embeddings, making them harder to classify.

Experimental Setup:

• Data: The Sonos Voice Control Bias Assessment Dataset, which provides granular metadata for gender, age, dialect, and ethnicity. The authors carefully partitioned SGs to isolate single variables and avoid confounding factors (e.g., ensuring gender tests aren't accidentally measuring ethnicity).
• Models: They evaluated several state-of-the-art Self-Supervised Speech Processing Models (S3Ms), including WavLM (base and large), Wav2vec 2.0, DeCoAR 2.0, and the end-to-end trained Whisper-medium.
• Probing (Measuring Bias): They used Phoneme Recognition (PR) probes (linear classifiers). They compared performance when probes were trained on a balanced dataset versus when they were trained on a single SG (in-domain training). If in-domain training significantly improved performance for a specific SG, it suggested the presence of embedding bias.
• KNN Distance (Measuring Variance): They directly calculated the k-nearest neighbors (KNN) distance between embeddings of the same phoneme for individual speakers. Higher KNN distances indicate higher embedding variance (random error).
• Fairness Intervention: They tested whether Domain Enhancing/Adversarial Training (DET/DAT)—a common technique to minimize speaker information in later layers—actually mitigates these errors.

Key Contributions

1. Error Typification Framework: A novel way to categorize ASR unfairness into systematic bias vs. random variance.
2. Identification of the Primary Culprit: The study demonstrates that while embedding bias exists in early layers, unequal variance (random error) is the more prevalent and persistent cause of unfairness in the layers used for ASR.
3. Evaluation of Fairness Algorithms: The paper provides empirical evidence that current fairness-enhancing algorithms (DET/DAT) are largely ineffective at addressing phoneme-level unfairness because they target bias rather than variance.

Results

• Bias vs. Variance: In-domain probe training provided only marginal benefits for some SGs (like dialect), suggesting that embedding bias is relatively low in the later layers of ASR-finetuned models. Conversely, the SGs that performed worst on PR tasks were almost always the same SGs that exhibited the highest KNN distances (highest variance).
• Correlation: There is a statistically significant correlation between high KNN distance and poor phoneme prediction accuracy across all tested models and SGs.
• Ineffectiveness of DET/DAT: Finetuning with DET/DAT did not significantly reduce the relative KNN distance between SGs, nor did it improve the benefits of in-domain training. This suggests that the "unfairness" is baked into the variance of the representations, which adversarial training does not resolve.
• Demographic Trends: Children and non-native speakers (Latino/Asian dialects) consistently showed higher embedding variance, making them harder for the models to represent accurately.

Significance

This research shifts the focus of ASR fairness research from "correcting shifts in representation" (bias) to "stabilizing the quality of representation" (variance). It suggests that to build truly fair ASR systems, researchers should move away from adversarial training aimed at erasing speaker identity and instead explore methods to reduce embedding dispersion for disadvantaged groups—such as supervised contrastive learning or Siamese networks—to ensure that phoneme representations are as precise and stable for all speakers as they are for the majority groups.