The Open Platform for AI Music Therapy Research

NeuroTunes generates personalized, physiologically-informed therapeutic music — and gives clinicians and ML researchers the tools to study it, improve it, and build on it.

NeuroTunes™ — Patent Pending · Open source, closed weights

Research Platform — Educational & Research Use Only — Not a Medical DevicePatent PendingOpen source · closed weights

Choose your interface

The same platform, three tailored entry points.

For Users

Generate personalized therapeutic music from how you feel right now, listen, and rate what helps. Simple, guided, and private. Anyone is welcome to try the research experience — it is not a medical device or treatment.

For Clinicians & Music Therapists

Clinical controls with an IRB / informed-consent panel. Record a protocol reference in the audit trail and generate consent-verified sessions for studies. In this demo the IRB reference is optional; a production research deployment requires it.

For ML Researchers

A programmatic API console: authenticate with an API key, generate music, export de-identified RLHF training data, explore the interactive OpenAPI reference, and inspect the federated network in the Research Observatory.

What the platform provides

Everything below is implemented in the open-source code today.

Physiologically-informed generation

Emotion → clinical-parameter mapping with binaural-beat encoding, bounded by an explicit parameter safety validator.

Periodic RLHF

Feedback-driven reward model with held-out R² + k-fold cross-validation and champion/challenger promotion gating.

Federated architecture

Federated coordinator with secure aggregation (HE + Shamir) and differential privacy (ε = 1.0) for multi-site training.

Consent-aware by design

A basic IRB / informed-consent gate (verify_consent) records a protocol reference alongside consent-gated sessions.

Versioned public API

Scoped, rate-limited API keys over a stable /api/v1 surface — built for machine-to-machine research use.

Open & reproducible

Open-source code, an OpenAPI spec, interactive Swagger docs, and a Python SDK. Model weights are closed.

Roadmap & status

An honest snapshot of what is built versus what is future work. We do not claim clinical-grade completeness that has not been validated.

Done Partial Planned
Periodic RLHF retrainingDone

Weekly batch retraining with a 50+ feedback threshold and t-test significance gating.

Reward model R² + k-fold CVDone

Held-out R²/MSE and 5-fold cross-validation persisted to reward_metrics.json.

Champion / challenger promotionDone

A challenger is promoted only if its held-out R² ≥ champion − tolerance, else rollback.

Parameter safety boundariesDone

Hard clamps + recommended-range warnings wired into both generation endpoints.

Federated coordinator + secure aggregationDone

Coordinator, HE + Shamir secure aggregation, and differential privacy (ε = 1.0).

Model versioning + distributed rollbackDone

Version registry with manual/approval-gated distributed rollback and risk assessment.

Treatment-outcome trackingPartial

Clinical efficacy scoring exists; a longitudinal per-user outcome dashboard does not yet.

Condition-specific adapters + embeddingsPlanned

The flagship open ML item: shared backbone + per-condition adapters for cross-condition learning.

Multi-site empirical federated validationPlanned

Infrastructure is built; validation across ≥2 real sites is future work with partner institutions.

Edge / on-device sentiment analysisPlanned

A privacy feature (client-side sentiment); not present today.