Exaia AI Research Lab
At Exaia, we’re not just embracing the future of AI — we’re shaping it with Human-Centered and Explainable AI (HC-XAI) research at our core.
Exaia’s team conducts rigorous scientific research and couples it with deep software engineering expertise to turn validated evidence into deployable technology. Curated datasets, robust metrics, empirical validation, and benchmarking are embedded throughout our end-to-end development pipeline.
100+
Peer-reviewed publications
3B+
Language tokens processed
200M+
Curated data points analyzed
1,000+
Models trained and benchmarked
Curated by our research team
Our team designs studies and builds research datasets spanning experimental behavioral measures, context-rich speech and language corpora, clinical cohorts, and longitudinal digital phenotyping.
01/Cognition & Behavior
Fine-grained behavioral records linked to experimentally assessed cognitive capacities, collected through controlled laboratory studies and rigorously designed crowdsourcing protocols.
02/Individual Differences
Context-rich communication corpora linking patterns of expression to individual profiles, established trait measures, and experimentally assessed psychological constructs.
03/Social Communication
Developmental and communication resources for investigating how social abilities emerge, how communicative intent is expressed, and how authenticity and veracity vary across contexts.
04/Brain & Mental Health
Clinically characterized neurological and psychiatric cohorts complemented by large-scale, naturalistic resources for studying cognitive and behavioral variation over time.
Science informs technology
Forged at our alma mater RWTH Aachen University, where our three co-founders met and connected siloed avenues across Computational Linguistics, Cognitive Sciences, and Responsible AI, coupling them with Software Engineering and Privacy-Preserving Technologies.
We have internalized rigorous metrics, empirical validation, and benchmarking—ensuring strong ties between science and deployable technology.
01
Empirical evaluation connects model performance with meaningful cognitive and clinical questions.
02
Published evaluations on international shared tasks, including SMM4H and ADReSSo.
03
Maintainable software and privacy-preserving design translate research into dependable systems.
Accuracy–interpretability trade-off
Model credibility demands more than predictive accuracy. In healthtech, we benchmark established performance metrics—accuracy, F1, precision and recall—and systematically evaluate interpretability, computational efficiency, privacy and robustness. Scientific and clinical validation test whether predictions reflect meaningful signals, generalize to the intended population and support the intended clinical use. Together, these assessments guide model selection and deployment.
Pareto analysis
Model classes positioned by predictive performance and interpretability. Marker size indicates computational cost.
Decision trail · each model benchmarked against the best interpretable option so far
Expert knowledge
Expert-engineered acoustic and linguistic biomarkers with directly readable weights: the transparent starting point for every task.
Beyond accuracy · clinical deployment profile
Explained through
The interpretable baseline every other model is benchmarked against. The adopted Hybrid fusion trades some transparency for a justified performance gain.
Our expertise
Balancing accuracy and interpretability is not a one-off decision. It is our ongoing practice.
We have run extensive experiments across clinical tasks to know when fully interpretable models are sufficient and when higher-dimensional representations earn a justified gain. As data, tasks and architectures evolve, we continue to train, update and deploy models, and re-test the trade-off every time.
Schematic illustration of the trade-off we evaluate. Positions are conceptual, not measured results.
Our 2023 study compares BiLSTMs, MentalRoBERTa and information-infused multitask models, with ablation, LIME and AGRAD analyses.
Read our research ↗International competitions & research benchmarks
We repeatedly benchmark our AI models in international shared-task competitions and on established research datasets—from Alzheimer’s disease and mental-health detection to personality prediction and emotion recognition. Evaluation under common conditions makes performance directly comparable with competing approaches, reveals strengths and limitations, and provides a rigorous basis for model selection and improvement.
Organizing institutions
Organizing institutions include
Organizing institutions include
PIM-BrainTwin
The brain is a multiscale system. To understand it, we need a new approach that brings together research still pursued in isolation, across scales, methods, and disciplines.
With partners at Imperial College London and OTH Regensburg, we are advancing that work on the drivers of brain ageing and neurodegeneration—and the considerable heterogeneity in individual disease trajectories.
Understanding why individuals differ in vulnerability, progression, and treatment response requires connecting molecular, imaging, clinical, environmental, and behavioral information across biological scales and time. Our speech biomarker technologies contribute repeatable measures of cognitive and behavioral function to this broader effort, linking everyday functional performance with patient-specific models of underlying disease processes.
Precision digital twins provide a framework for integrating these complementary observations into models that evolve as new information becomes available. Our research explores how this integration can support mechanistic understanding, individual trajectory modelling, and the investigation of intervention responses.
Our PIM-BrainTwin perspective develops one example: a physics-informed, multiscale framework for estimating underlying disease states and remaining compensatory capacity. Alzheimer’s disease serves as the initial use case within an open, modular research framework designed to accommodate other neurological conditions and alternative mechanistic hypotheses.
Research in the literature
Selected work in speech biomarkers, mental health and responsible AI.
D Wiechmann, E Kerz, M Albrecht, Y Qiao, J Pinho, K Reetz, AS Costa
Frontiers in Neurology 17, 1780783
View publicationD Wiechmann, A Günes, E Kerz, Y Qiao, M Köhne, U Sprick
Journal of Alzheimer’s Disease, 13872877261456294
View publicationA Nonn, E Kerz, D Wiechmann, P Edison
arXiv preprint arXiv:2608.15342
View publicationEK Saad Mankarious, Y Qiao, D Wiechmann, E Kerz, A Zirikly
Proceedings of the Fifteenth Language Resources and Evaluation Conference …
View publicationD Wiechmann, E Kerz, E Kempa, Y Qiao
Proceedings of the Fifteenth Language Resources and Evaluation Conference …
View publicationE Kerz, S Zanwar, Y Qiao, D Wiechmann
Frontiers in psychiatry 14, 1219479
View publicationOur research connects collaborators across universities and clinical institutions, bringing together expertise in speech and language science, responsible AI, neuroscience, clinical research, and computational modelling.
Institutions represented in our research collaborations
We welcome researchers, clinicians, and industry partners who share our ambition to turn scientific insight into trustworthy technologies for brain health. Collaborate with us on speech biomarker development, multimodal modelling, longitudinal studies, clinical validation, and precision digital twin frameworks.
Discuss a research collaboration