Real-World Evaluation of a Consumer Voice-Based AI Emotional Support System: CORE-10 Outcomes and Usage Associations

Background: Voice-first generative-AI agents are increasingly deployed for emotional support, yet real-world outcome evidence remains limited: standardized reliable-change and recovery metrics are largely absent from this literature and, to our knowledge, unreported for any voice-first consumer generative-AI app. Objective: To evaluate a deployed consumer voice-based AI emotional-support system (Lovon) in real-world use: longitudinal CORE-10 outcomes (reliable change and recovery), usage-outcome associations, and exploratory within-session emotional-state shifts. Methods: Retrospective observational analysis of routinely collected product data (February-July 2026). 5,882 adult users completed 6,904 in-app CORE-10 assessments; 656 completed two or more (paired sample); the 14-day-or-longer interval was the focal horizon, with all-paired results reported alongside. Session-level emotional state (0-10) was scored by a large language model; 31,277 sessions were analyzable after excluding pipeline-failure defaults. Results: At first assessment, 93.7% of users scored at or above the clinical cutoff (mean 23.0); 45.5% began in the severe range. Across all paired users (n=656), 38.9% showed reliable improvement (d=0.57). In the focal subgroup with 14 or more days between assessments (n=257, median 26 days), reliable improvement was 46.3% (95% CI 40.3-52.4) versus 5.4% deterioration (8.6:1), d=0.76; 22.9% recovered. Estimates were larger at longer horizons (50.0%, d=0.85 at 56+ days) and with higher usage: the top tercile reached 60.0% improvement and 45.9% recovery (d=1.14) versus 41.2% and 0% among minimal users, despite starting less severe. 78.5% of sessions of 10+ minutes beginning in distress ended improved (mean +2.0). Safety audits flagged 2.2% of 45,111 sessions (system activity, not validated detection). Conclusions: Observational associations in self-selected completers; no causal claims. Findings support feasibility and motivate prospective controlled evaluation. Reported to STROBE and RECORD guidance; analysis code: doi.org/10.17605/OSF.IO/MCDJZ. This preprint has not been peer reviewed.