Responsible AI

As speech and language technologies become more capable, ensuring that they are developed and deployed responsibly becomes increasingly important. AI systems that recognise speech, generate voices, and interact with people raise questions that extend beyond technical performance, including fairness, consent, data governance, and the consequences of system failures. The Responsible AI team investigates these challenges across Africa’s diverse linguistic and social contexts, examining where systems fail, who bears the consequences, how data is used, and what safeguards are needed. Its research considers responsibility throughout the AI lifecycle, from data sourcing and model development to evaluation, release, and deployment.

Research teams: Speech Intelligence · Language Intelligence · Language Resources · Model Intelligence · Responsible AI

Research Areas

Evaluation and fairness

AI systems do not perform equally across languages, dialects, accents and recording conditions. Aggregate results can hide failures that matter to particular communities, while evaluation datasets may inadequately represent the people a system is intended to serve.

Read more +

We investigate how speech and language models should be evaluated across their intended populations and operating environments. This includes language-specific benchmarks, performance disparities, dataset limitations, robustness testing and transparent reporting of results.

Our research is guided by the principle that model performance should be assessed in context. Evaluation should inform decisions about a system's intended uses, known limitations and readiness for release.

Voice safety and misuse prevention

Speech synthesis and voice cloning introduce risks that differ from those of conventional text-based systems. Generated speech can be used to impersonate people, mislead listeners or facilitate fraud. These risks depend on the capabilities of a system, how access is provided and the context in which its outputs are used.

Read more +

We examine the risks associated with voice generation and voice-based interaction, including speaker consent, impersonation, access controls, misuse prevention and the identification of synthetic speech where technically feasible.

Our work considers safeguards across the lifecycle of voice technologies, from decisions about model capabilities and access to monitoring potential misuse and responding to newly identified risks. We investigate how safeguards should be proportionate to a system's capabilities and foreseeable harms, recognising that no single technical measure can eliminate every form of misuse.

Privacy and data governance

Speech and language research depends on data, but the availability of a recording or text does not automatically establish the right to use it. Voices can reveal identity and other personal information, while datasets may carry restrictions arising from consent, licensing, provenance or the circumstances in which they were collected.

Read more +

We investigate the principles and technical practices that should govern data collection, provenance, consent, permitted uses, retention, access and responsible data sharing. Particular attention is given to recorded voices, speaker-identifying information and derived representations that may retain sensitive characteristics. These questions connect directly to our privacy commitments.

We consider governance throughout the research lifecycle: before data is accepted, while it is processed and used in model development, and when datasets, models or research outputs are shared. This includes examining what should be documented, which uses are permitted, and what restrictions or safeguards may be necessary. Our aim is to connect responsible data practices with reproducible research without treating openness as a substitute for permission or accountability.

Language communities and context

Responsible AI must account for the people whose languages, voices and knowledge shape the systems being built. Languages are not interchangeable, and linguistic differences extend beyond vocabulary to pronunciation, dialect, cultural meaning and patterns of use. Systems developed without sufficient contextual understanding may exclude communities, misrepresent speech or perform poorly in settings where reliability matters.

Read more +

We explore ways to incorporate native-speaker expertise, linguistic knowledge and community perspectives into research design, data development, evaluation and documentation. This includes examining how language-specific needs should influence the selection of research priorities, the interpretation of evaluation results and decisions about how models are presented and used.

Our guiding principle is that technical development should be informed by the contexts in which systems are intended to operate. We investigate how community perspectives can meaningfully contribute to research decisions and how limitations, intended uses and potential consequences can be communicated to the people affected by these technologies.

Human oversight and responsible deployment

AI systems can produce convincing outputs without being reliable in every situation. In applications such as voice agents, failures may affect people's understanding, decisions or access to services. Responsible deployment therefore requires more than demonstrating that a model works under controlled conditions.

Read more +

We examine appropriate boundaries for automation, human oversight, escalation and fallback, alongside the conditions under which a system may need to be restricted, corrected or withdrawn. Our research considers the relationship between model capabilities, the consequences of failure and the level of human involvement appropriate for a given application.

We approach deployment as a lifecycle rather than a single release event. This means investigating how evaluation findings, known limitations, intended-use boundaries and emerging evidence of failure should inform release decisions and subsequent changes. We also consider the governance needed to assign responsibility, document important decisions and respond when a system no longer meets its intended requirements.

Our goal is to connect responsible AI principles with practical engineering decisions.