Automatic speech recognition (ASR) still struggles with pathological speech, according to a new preprint on arXiv. The core issue is a persistent mismatch between the characteristics of a given speaker's voice and the models ASR systems are trained on. In clinical and accessibility contexts, this mismatch is especially challenging because only a small amount of the target speaker's data is typically available for adaptation.
The paper, titled "ASR Personalization in Pathological Voice: Accuracy, Efficiency, and Duration Effects," appears to focus on how personalization techniques can be applied under these data constraints. The abstract highlights accuracy, efficiency, and duration as key dimensions of the problem, but the full findings are not yet available from the abstract alone. As a preprint, the work has not yet undergone peer review.