AI for Speech-Language Pathologists

SLPs use AI in three main ways, plus paperwork. The first is making materials. Language models quickly produce word lists, sentences, stories, social scenarios and conversation prompts at a chosen level and on a chosen theme. The catch is that models work with written text, not sounds. Asked for words that start with /k/, a model may include 'knife' or 'city' because they begin with the letters k or c. Asked for /f/ words, it may leave out 'phone.' The clinician has to check every list phonetically, and also check complexity: syllable count, where the sound falls in the word, and blends. The second is speech analysis. Automatic speech recognition (ASR) can give a first-draft transcript of a language sample, and some tools estimate measures like speaking rate or pauses. But ASR systems are trained mostly on typical adult speech. Their error rates rise sharply for children and for people with speech sound disorders, dysarthria, stuttering or some accents. ASR also tends to 'correct' errors into the word the speaker meant. A child who says 'wabbit' may be transcribed as 'rabbit', which hides the very error you are measuring. Google's Project Relate and the Speech Accessibility Project at the University of Illinois are trying to improve recognition of atypical speech by training on it. The third is practice tools. Articulation apps offer drills and recording, and some give automatic feedback on whether the target sound was correct. Automatic feedback can increase practice between sessions, which matters because many repetitions support motor learning. But its accuracy varies, and a wrong 'correct' signal can reinforce errors. Test an app's feedback against your own ears before you assign it. AI also helps with paperwork: drafting evaluation reports, wording IEP goals and writing summaries for parents. ASHA's Code of Ethics still requires competence and protection of client information, and that covers anything you upload. A common misconception is that an app's accuracy score is the same as a clinical judgment. It is one noisy data point.