Virtual Speech Therapist: A Clinician-in-the-Loop AI Speech Therapy Agent for Personalized and Supervised Therapy

A team of researchers has introduced Virtual Speech Therapist (VST), an AI-driven platform designed to automate stuttering assessment and generate personalized therapy plans while keeping clinicians in control of final decisions, according to a paper posted on the arXiv preprint server [1]. The system, detailed in a manuscript submitted on 1 May 2026 and revised on 15 June 2026, combines deep-learning models that classify stuttering types with multi-agent large language model (LLM) reasoning [1]. After a patient’s speech sample is acquired and analyzed, specialized LLM agents autonomously propose, critique, and refine an individualized therapy draft [1]. A separate critic agent then evaluates each plan for clinical safety, methodological soundness, and alignment with peer-reviewed evidence and professional guidelines [1]. The final output is a draft intended for a speech-language pathologist to review and modify, a workflow the authors describe as a clinician-in-the-loop paradigm [1]. Expert speech therapists who tested the platform confirmed that VST consistently produces high-quality, evidence-based therapy recommendations [1]. The authors argue the tool could reduce the documentation and planning burden on clinicians while improving outcomes for people who stutter [1]. An interactive interface for the system is available online, allowing real-time stuttering assessment and customized therapy planning [1]. The paper appeared on arXiv, an open-access repository that hosts electronic preprints across disciplines including computer science, physics, and mathematics [6]. Submissions to arXiv are moderated but not peer-reviewed, and the repository has grown to a rate of roughly 24,000 new articles per month as of late 2024 [6]. The VST manuscript is listed under the artificial intelligence category and is accompanied by experimental browser tools developed through arXivLabs, a framework that lets community collaborators build features directly on the site [4][5]. Large language models, the technology underpinning VST’s reasoning agents, are machine-learning systems with many parameters trained on vast text corpora for tasks such as language generation [8]. The VST platform applies that capability to a clinical domain, using multiple agents that check one another’s outputs before a human therapist ever sees the plan [1]. The paper does not report clinical trial data, and the authors frame the work as a demonstration of the system’s potential to augment—not replace—human clinical judgment [1].