Researchers have built a proof-of-concept system called the Virtual Speech Therapist (VST) that aims to support stuttering assessment and therapy planning. The system pairs deep-learning-based stuttering detection with a team of large language model (LLM) agents that generate, critique, and refine individualized therapy plans. The work is described in a preprint on arXiv.

In the proposed workflow, speech samples are first processed to detect stuttering types. Those speech-driven signals then feed into an agentic reasoning process where specialized LLM agents produce a draft therapy plan. A dedicated CriticAgent reviews the draft for consistency, methodological soundness, and alignment with professional guidelines, iteratively improving it before presentation to a clinician. The clinician can then give feedback, and the system revises the plan accordingly, maintaining a clinician-in-the-loop paradigm.

An initial evaluation with expert speech-language pathologists examined the clinical interpretability and feasibility of the generated plans. The results suggest the system could augment clinician-led therapy planning, though the authors note this is only a proof of concept. They call for future studies to evaluate integration into real-world clinical settings, and an interactive prototype is available online. The paper is under review.```json {