Research

The work asks how models can learn useful representations from noisy, variable and multimodal data while remaining open to inspection. Speech and affect provide demanding evaluation settings. Current research directions extend toward reliable physiological and behavioral signal modeling, uncertainty and visual model verification.

The bureau connects each research direction to its papers and software. These records distinguish published evidence from future directions and preserve the conditions under which results were obtained.

Deep learning, learnable representations and time-frequency analysis; multimodal learning; explainable and responsible AI. Evaluation interests include calibration, robustness, uncertainty, experimental design and statistical testing.

A recurring question is how a model can preserve useful mathematical structure while adapting to a task—and how its predictions can remain inspectable by a domain expert.

Speech and audio, behavioral and physiological signals, time series, and multimodal combinations of text, audio and video. The research addresses noisy, incomplete and sensitive data, including crisis-call recordings handled within an approved ethics framework.

Speech emotion recognition is an evaluation setting for methodological work. Extending reliable modeling and visual model verification to health is a research direction, with the requirements of each application kept explicit.

Python, PyTorch, TensorFlow, R and SQL support the modeling and analysis work; Git, Docker and Linux support software development and reproducibility. Public research repositories connect implementations with papers, configurations and evaluation instructions.

The research process combines software tests and version control with ablations, error analysis, robustness checks and attention to negative results. Reproducibility includes documenting data-access conditions as well as making code available.