Client speech recognition techniques don’t work as nicely for many individuals with speech variations, similar to stuttering, relative to the remainder of the final inhabitants. Nevertheless, what just isn’t clear is the diploma to which these techniques don’t work, how they are often improved, or how a lot folks wish to use them. On this paper, we first deal with these questions utilizing outcomes from a 61-person survey from individuals who stutter and discover contributors wish to use speech recognition however are continuously reduce off, misunderstood, or speech predictions don’t symbolize intent. In a second examine, the place 91 individuals who stutter recorded voice assistant instructions and dictation, we quantify how dysfluencies impede efficiency in a consumer-grade speech recognition system. By three technical investigations, we exhibit what number of widespread errors will be prevented, leading to a system that cuts utterances off 79.1% much less usually and improves phrase error charge from 25.4% to 9.9%.
