Aphasia is a language disorder due to brain damage, often after a stroke. Those affected have difficulty speaking, understanding, reading, or writing. Aphasia manifests differently in each individual, despite the presence of recognizable patterns of language impairment associated with specific brain lesion locations.
In clinical practice, aphasia is usually diagnosed using standardized language tests. These procedures have proven effective, but they are time-consuming and heavily dependent on human expertise. This is because speech samples must be listened to, transcribed, and contextualized by specialists.
Given the rising number of cases and limited human resources, this poses a growing challenge for everyday clinical practice. Automated approaches can provide targeted relief here without replacing professional decisions.
Even powerful, generally trained speech-based AI models quickly reach their limits if they are not specifically prepared for the peculiarities of aphasic speech. This is precisely where a study involving audEERING comes in. It was developed as part of the autoAAT consortium funded by the BMBF (now BMFTR) and investigates how AI models can automatically analyze aphasic speech and provide clinical support. The study is based on retrospectively evaluated speech data from routine clinical practice, which was collected over several years as part of the Aachen Aphasia Test (AAT), the standard diagnostic procedure in German-speaking countries (source: Huber, W., Poeck, K., Willmes, K.: The aachen aphasia test. In: Rose, C. (ed.) Advances in Neurology vol. 42, pp. 291–303. Raven Press, New York (1984)).
AI models trained with aphasia data deliver better results
The results show that AI models deliver significantly better results when they are specifically prepared with aphasia data (domain adaptation) and analyze answers not in isolation but in a question-answer context. The study examines whether the model achieves better results when it knows not only the answer but also the corresponding question. This is particularly relevant because short or incomplete answers in aphasia can only be correctly classified if one knows what they refer to.
The most powerful model achieves a correlation of 0.66 (on a scale of 0 to 1) between the automatic analysis and the clinical assessment. The result shows that changes in speech status recorded by specialists in the test are reliably reproduced by the system in most cases.
The result is a system that enables continuous monitoring of the progression of aphasia and moves away from the classification of aphasia syndromes, which tend to remain stable over time and therefore offer only limited insights into the course of the disease.
Robust automatic assessment of speech symptoms can also help clinicians evaluate recovery progress, compare therapy outcomes, and optimize treatment parameters such as therapy approach, dosage, frequency, and intensity in both clinical and research contexts. As a result, a multicenter study is needed to further evaluate the applicability of our work.
I am very excited to present these exciting results at the Voice AI Symposium & Hackathon from May 4 to 6 in St. Petersburg, Florida!
