Development and validation of a voice-based conversational artificial intelligence coaching agent to reduce competitive state anxiety in elite athletes: study protocol for a randomized, analyst-blinded, attention-matched active-controlled trial

Background Competitive state anxiety harms elite athletes’ performance and mental health, while stigma and training demands limit access to face-to-face support; voice-based conversational artificial intelligence (AI) analyzing speech in real time offers a scalable route to psychological skills training (PST). Methods Phase I develops the agent and tests content validity and usability against prespecified thresholds. Phase II is a parallel-group, analyst-blinded, attention-matched active-controlled randomized trial, fully online. All outcomes are self-reported by unblinded participants, so no outcome assessor is blinded; blinding covers the data manager and statistician. Eighty-six elite athletes registered with the Korean Sport and Olympic Committee are randomized 1:1 to 4 weeks of voice-based AI coaching (three sessions/week, ≥20 min) or a wordless relaxation recording, with delayed access. The arms match on dose, frequency and digital delivery but not on interactivity, content variety, platform or expectancy, so the trial estimates the incremental effect of the complete agent over a non-interactive comparator rather than isolating voice adaptation; credibility, expectancy, and adherence are measured and modeled in both arms. The primary endpoint is change from baseline to week 4 in Competitive State Anxiety Inventory-2 cognitive anxiety, administered within 24 h before the index competition nominated for each window; somatic anxiety and self-confidence follow in a fixed sequence. Secondary outcomes are GAD-7, PHQ-9, PSS-10, WHO-5 and SF-12; SUS and an AI voice-based anxiety index are exploratory. Timepoints are fixed relative to the first intervention session: baseline (before randomization), within 7 days after the intervention ends, and 2 weeks later; prespecified contingencies govern a canceled or postponed competition. Analysis is by linear mixed models (intention-to-treat). Discussion This may be among the first randomized trials of a conversational AI agent analyzing athletes’ speech in real time to deliver tailored PST. Strengths are phase I development, a dose-matched active comparator and a decentralized design, although the comparator does not isolate voice adaptation. Powered for d = 0.60 on the primary contrast, it is as much estimation as hypothesis testing. Clinical Trial Registration https://cris.nih.go.kr , identifier KCT0012076.