Data powering AI is often opaque. Researchers, NGOs, and law and policy leaders have called for greater transparency about how data is used for training, fine-tuning, and evaluation. While data transparency is often championed as crucial, what it concretely enables is largely implicit. Similarly, the concerns developers seem to have about transparency go unstated. This lack of clarity has led some researchers to critique transparency demands as disconnected from the actual benefits—or risks—to specific stakeholders. We analyze documentation from four stakeholder groups to create a taxonomy of the risks and benefits of dataset transparency. Data transparency is perceived as either a risk or a benefit given a stakeholder’s position, rather than wholesale. We also propose data availability and data documentation as two lenses through which to consider transparency. We discuss how best to strategically promote situational data transparency that takes into account the relationship between stakeholder position, transparency modality, and benefits/risks.

Speaker recognition is a widely used voice-based biometric technology with applications in various industries, including banking, education, recruitment, immigration, law enforcement, healthcare, and well-being. However, while dataset evaluations and audits have improved data practices in face recognition and other computer vision tasks, the data practices in speaker recognition have gone largely unquestioned. Our research aims to address this gap by exploring how dataset usage has evolved over time and what implications this has on bias, fairness and privacy in speaker recognition systems. Previous studies have demonstrated the presence of historical, representation, and measurement biases in popular speaker recognition benchmarks. In this paper, we present a longitudinal study of speaker recognition datasets used for training and evaluation from 2012 to 2021. We survey close to 700 papers to investigate community adoption of datasets and changes in usage over a crucial time period where speaker recognition approaches transitioned to the widespread adoption of deep neural networks. Our study identifies the most commonly used datasets in the field, examines their usage patterns, and assesses their attributes that affect bias, fairness, and other ethical concerns. Our findings suggest areas for further research on the ethics and fairness of speaker recognition technology.

Billions of distributed, heterogeneous, and resource constrained IoT devices deploy on-device machine learning (ML) for private, fast, and offline inference on personal data. On-device ML is highly context dependent and sensitive to user, usage, hardware, and environment attributes. This sensitivity and the propensity toward bias in ML makes it important to study bias in on-device settings. Our study is one of the first investigations of bias in this emerging domain and lays important foundations for building fairer on-device ML. We apply a software engineering lens, investigating the propagation of bias through design choices in on-device ML workflows. We first identifyreliability biasas a source of unfairness and propose a measure to quantify it. We then conduct empirical experiments for a keyword spotting task to show how complex and interacting technical design choices amplify and propagatereliability bias. Our results validate that design choices made during model training, like the sample rate and input feature type, and choices made to optimize models, like light-weight architectures, the pruning learning rate, and pruning sparsity, can result in disparate predictive performance across male and female groups. Based on our findings, we suggest low effort strategies for engineers to mitigate bias in on-device ML.

Speaker verification (SV) provides billions of voice-enabled devices with access control, and ensures the security of voice-driven technologies. As a type of biometrics, it is necessary that SV is unbiased, with consistent and reliable performance across speakers irrespective of their demographic, social and economic attributes. Current SV evaluation practices are insufficient for evaluating bias: they are over-simplified and aggregate users, not representative of usage scenarios encountered in deployment, and consequences of errors are not accounted for. This paper proposes design guidelines for constructing SV evaluation datasets that address these short-comings. We propose a schema for grading the difficulty of utterance pairs, and present an algorithm for generating inclusive SV datasets. We empirically validate our proposed method in a set of experiments on the VoxCeleb1 dataset. Our results confirm that the count of utterance pairs/speaker, and the difficulty grading of utterance pairs have a significant effect on evaluation performance and variability. Our work contributes to the development of SV evaluation practices that are inclusive and fair.

When deploying machine learning (ML) models on embedded and IoT devices, performance encompasses more than an accuracy metric: inference latency, energy consumption, and model fairness are necessary to ensure reliable performance under heterogeneous and resource-constrained operating conditions. To this end, prior research has studied model-centric approaches, such as tuning the hyperparameters of the model during training and later applying model compression techniques to tailor the model to the resource needs of an embedded device. In this paper, we take a data-centric view of embedded ML and study the role that pre-processing parameters in the data pipeline can play in balancing the various performance metrics of an embedded ML system. Through an in-depth case study with audio-based keyword spotting (KWS) models, we show that pre-processing parameter tuning is a remarkable tool that model developers can adopt to trade-off between a model's accuracy, fairness, and system efficiency, as well as to make an embedded ML model resilient to unseen deployment conditions.