NeuroTunes: An Open, Reproducible Platform for Clinicians and Researchers to Build and Audit Therapeutic Music Generation

Evaluating therapeutic music-generation systems before human trials requires rigorous and reproducible pipelines. We demonstrate NeuroTunes, an open-source platform embedding evaluation directly into pre-clinical development. We serve three user groups through a unified consent-gated request path: patients submitting assessments for rendered tracks, clinicians running protocol-referenced sessions via a de-identified console selecting indication-based binaural parameters, and researchers operating pipelines via versioned REST APIs and Python SDKs. The renderer functions deterministically without learned audio models, and annotation passes use fixed seeds to guarantee cross-laboratory script reproducibility. Internally, EmotionNet maps assessments to arousal, valence, focus, and calm; parametric renderers convert vectors to audio; safety validators limit tempo, frequency, and duration. Evaluation operates continuously via confidence gates routing uncertain labels to human review, and an independent MERT-based audio critic (Music2Emo) unexposed during training evaluates rendered outputs within promote-and-rollback paths. NeuroTunes is released strictly as a research and educational tool making no claims of clinical efficacy.