Speaker: Unax Murua Urizarbarrena.
Abstract: This presentation covers two different projects using data and speech analysis. The first half focuses on building a machine learning model to predict sleep apnea within the INSPIRA project (Colaboration Network funded by the Comunidad de Madrid regional government). Using a clinical dataset from the research institute of La Paz Hospital (IDIPaz), we look at how reducing the number of variables affects the model’s performance. By comparing newer AI models like TabPFN with traditional linear models and standard medical questionnaires, we show that carefully selecting just a few key variables can create a reliable scoring system that consistently works better than standard screening tools.
The second half shifts to the COSER project, a massive collection of over 2,000 hours of rural Spanish audio. Here, we focus on specific tasks designed both to fix old transcriptions and to help guide state-of-the-art diarization systems to improve their performance. The tasks themselves are detecting whether or not voices overlap, and identifying if there is only one speaker or multiple. Our main goal is to evaluate if Audio-LLMs can be a useful tool for solving these challenges, though we also test several other models for comparison.