Fairness Constraints in Speech-to-Text Applications – Clinical Effects of Salivation

Speech-to-text (StT) applications tend to become ubiquitous as AI interfaces with devices. This raises medical, technical, and ethical concerns regarding the ability of users with various health conditions to interact with the StT interfaces. The tests reported in the paper bring evidence supporting the concerns and indicate some of the issues. 18 subjects under neuroleptic drugs were investigated for the dentition condition, salivation (with a sialometer), and speech performance, specifically for pronunciation of words with fricatives and of isolated fricatives. The recognition rate by an StT application was tested. Contrary to what is known for healthy subjects, for some subjects with neurological conditions and under antipsychotics, stimulated salivation can decrease speech intelligibility and reduce efficiency of interacting with devices using speech recognition. Clinical relevance of the results: doctors should not recommend saliva stimulating drugs without an in-person consult and before the effect on speech intelligibility by machines is assessed, because the effect may be detrimental. Significance: This is the first study that indicates a fairness issue of StT applications, in that for users with salivation conditions such as during treatment with medications that stimulate salivation, the StT tool may disadvantage them and have reduced usability for them.