Beyond Word Error Rate: Auditing the Diversity Tax in Speech Recognition through Dataset Cartography

Automatic speech recognition (ASR) systems are predominantly evaluated using the Word Error Rate (WER). However, raw token-level metrics fail to capture semantic fidelity and routinely obscures the `diversity tax', the disproportionate burden on marginalized and atypical speaker due to systematic recognition failures. In this paper, we explore the limitations of relying solely on lexical counts by systematically evaluating a broader class of non-linear and semantic metrics. To enable rigorous model auditing, we introduce the sample difficulty index (SDI), a novel metric that quantifies how intrinsic demographic and acoustic factors drive model failure. By mapping SDI on data cartography, we demonstrate that metrics EmbER and SemDist expose hidden systemic biases and inter-model disagreements that WER ignores. Finally, our findings are the first steps towards a robust audit framework for prospective safety analysis, empowering developers to audit and mitigate ASR disparities prior to deployment.

Measuring What VLMs Don't Say: Validation Metrics Hide Clinical Terminology Erasure in Radiology Report Generation

Reliable deployment of Vision-Language Models (VLMs) in radiology requires validation metrics that go beyond surface-level text similarity to ensure clinical fidelity and demographic fairness. This paper investigates a critical blind spot in current model evaluation: the use of decoding strategies that lead to high aggregate token-overlap scores despite succumbing to template collapse, in which models generate only repetitive, safe generic text and omit clinical terminology. Unaddressed, this blind spot can lead to metric gaming, where models that perform well on benchmarks prove clinically uninformative. Instead, we advocate for lexical diversity measures to check model generations for clinical specificity. We introduce Clinical Association Displacement (CAD), a vocabulary-level framework that quantifies shifts in demographic-based word associations in generated reports. Weighted Association Erasure (WAE) aggregates these shifts to measure the clinical signal loss across demographic groups. We show that deterministic decoding produces high levels of semantic erasure, while stochastic sampling generates diverse outputs but risks introducing new bias, motivating a fundamental rethink of how "optimal" reporting is defined.

Intra-Fairness Dynamics: The Bias Spillover Effect in Targeted LLM Alignment

Conventional large language model (LLM) fairness alignment largely focuses on mitigating bias along single sensitive attributes, overlooking fairness as an inherently multidimensional and context-specific value. This approach risks creating systems that achieve narrow fairness metrics while exacerbating disparities along untargeted attributes, a phenomenon known as bias spillover. While extensively studied in machine learning, bias spillover remains critically underexplored in LLM alignment. In this work, we investigate how targeted gender alignment affects fairness across nine sensitive attributes in three state-of-the-art LLMs (Mistral 7B, Llama 3.1 8B, Qwen 2.5 7B). Using Direct Preference Optimization and the BBQ benchmark, we evaluate fairness under ambiguous and disambiguous contexts. Our findings reveal noticeable bias spillover: while aggregate results show improvements, context-aware analysis exposes significant degradations in ambiguous contexts, particularly for physical appearance ($p< 0.001$ across all models), sexual orientation, and disability status. We demonstrate that improving fairness along one attribute can inadvertently worsen disparities in others under uncertainty, highlighting the necessity of context-aware, multi-attribute fairness evaluation frameworks.

Who Does Your Algorithm Fail? Investigating Age and Ethnic Bias in the MAMA-MIA Dataset

Deep learning models aim to improve diagnostic workflows, but fairness evaluation remains underexplored beyond classification, e.g., in image segmentation. Unaddressed segmentation bias can lead to disparities in the quality of care for certain populations, potentially compounded across clinical decision points and amplified through iterative model development. Here, we audit the fairness of the automated segmentation labels provided in the breast cancer tumor segmentation dataset MAMA-MIA. We evaluate automated segmentation quality across age, ethnicity, and data source. Our analysis reveals an intrinsic age-related bias against younger patients that continues to persist even after controlling for confounding factors, such as data source. We hypothesize that this bias may be linked to physiological factors, a known challenge for both radiologists and automated systems. Finally, we show how aggregating data from multiple data sources influences site-specific ethnic biases, underscoring the necessity of investigating data at a granular level.

EmoTale: An Enacted Speech-emotion Dataset in Danish

While multiple emotional speech corpora exist for commonly spoken languages, there is a lack of functional datasets for smaller (spoken) languages, such as Danish. To our knowledge, Danish Emotional Speech (DES), published in 1997, is the only other database of Danish emotional speech. We present EmoTale; a corpus comprising Danish and English speech recordings with their associated enacted emotion annotations. We demonstrate the validity of the dataset by investigating and presenting its predictive power using speech emotion recognition (SER) models. We develop SER models for EmoTale and the reference datasets using self-supervised speech model (SSLM) embeddings and the openSMILE feature extractor. We find the embeddings superior to the hand-crafted features. The best model achieves an unweighted average recall (UAR) of 64.1% on the EmoTale corpus using leave-one-speaker-out cross-validation, comparable to the performance on DES.

Exploring Local Interpretable Model-Agnostic Explanations for Speech Emotion Recognition with Distribution-Shift

We introduce EmoLIME, a version of local interpretable model-agnostic explanations (LIME) for black-box Speech Emotion Recognition (SER) models. To the best of our knowledge, this is the first attempt to apply LIME in SER. EmoLIME generates high-level interpretable explanations and identifies which specific frequency ranges are most influential in determining emotional states. The approach aids in interpreting complex, high-dimensional embeddings such as those generated by end-to-end speech models. We evaluate EmoLIME, qualitatively, quantitatively, and statistically, across three emotional speech datasets, using classifiers trained on both hand-crafted acoustic features and Wav2Vec 2.0 embeddings. We find that EmoLIME exhibits stronger robustness across different models than across datasets with distribution shifts, highlighting its potential for more consistent explanations in SER tasks within a dataset.

Is it the model or the metric -- On robustness measures of deeplearning models

Determining the robustness of deep learning models is an established and ongoing challenge within automated decision-making systems. With the advent and success of techniques that enable advanced deep learning (DL), these models are being used in widespread applications, including high-stake ones like healthcare, education, border-control. Therefore, it is critical to understand the limitations of these models and predict their regions of failures, in order to create the necessary guardrails for their successful and safe deployment. In this work, we revisit robustness, specifically investigating the sufficiency of robust accuracy (RA), within the context of deepfake detection. We present robust ratio (RR) as a complementary metric, that can quantify the changes to the normalized or probability outcomes under input perturbation. We present a comparison of RA and RR and demonstrate that despite similar RA between models, the models show varying RR under different tolerance (perturbation) levels.

Examining the Interplay Between Privacy and Fairness for Speech Processing: A Review and Perspective

Speech technology has been increasingly deployed in various areas of daily life including sensitive domains such as healthcare and law enforcement. For these technologies to be effective, they must work reliably for all users while preserving individual privacy. Although tradeoffs between privacy and utility, as well as fairness and utility, have been extensively researched, the specific interplay between privacy and fairness in speech processing remains underexplored. This review and position paper offers an overview of emerging privacy-fairness tradeoffs throughout the entire machine learning lifecycle for speech processing. By drawing on well-established frameworks on fairness and privacy, we examine existing biases and sources of privacy harm that coexist during the development of speech processing models. We then highlight how corresponding privacy-enhancing technologies have the potential to inadvertently increase these biases and how bias mitigation strategies may conversely reduce privacy. By raising open questions, we advocate for a comprehensive evaluation of privacy-fairness tradeoffs for speech technology and the development of privacy-enhancing and fairness-aware algorithms in this domain.

Evaluation of Large Language Models: STEM education and Gender Stereotypes

Large Language Models (LLMs) have an increasing impact on our lives with use cases such as chatbots, study support, coding support, ideation, writing assistance, and more. Previous studies have revealed linguistic biases in pronouns used to describe professions or adjectives used to describe men vs women. These issues have to some degree been addressed in updated LLM versions, at least to pass existing tests. However, biases may still be present in the models, and repeated use of gender stereotypical language may reinforce the underlying assumptions and are therefore important to examine further. This paper investigates gender biases in LLMs in relation to educational choices through an open-ended, true to user-case experimental design and a quantitative analysis. We investigate the biases in the context of four different cultures, languages, and educational systems (English/US/UK, Danish/DK, Catalan/ES, and Hindi/IN) for ages ranging from 10 to 16 years, corresponding to important educational transition points in the different countries. We find that there are significant and large differences in the ratio of STEM to non-STEM suggested education paths provided by chatGPT when using typical girl vs boy names to prompt lists of suggested things to become. There are generally fewer STEM suggestions in the Danish, Spanish, and Indian context compared to the English. We also find subtle differences in the suggested professions, which we categorise and report.

Exploratory Evaluation of Speech Content Masking

Most recent speech privacy efforts have focused on anonymizing acoustic speaker attributes but there has not been as much research into protecting information from speech content. We introduce a toy problem that explores an emerging type of privacy called "content masking" which conceals selected words and phrases in speech. In our efforts to define this problem space, we evaluate an introductory baseline masking technique based on modifying sequences of discrete phone representations (phone codes) produced from a pre-trained vector-quantized variational autoencoder (VQ-VAE) and re-synthesized using WaveRNN. We investigate three different masking locations and three types of masking strategies: noise substitution, word deletion, and phone sequence reversal. Our work attempts to characterize how masking affects two downstream tasks: automatic speech recognition (ASR) and automatic speaker verification (ASV). We observe how the different masks types and locations impact these downstream tasks and discuss how these issues may influence privacy goals.