MeloTune: On-Device Arousal Learning and Peer-to-Peer Mood Coupling for Proactive Music Curation
A 94K-parameter iPhone model predicts your mood trajectory and syncs it with friends — no cloud, no raw data leaving the device.
The Thesis
MeloTune demonstrates that affective computing — real-time mood inference, per-user arousal personalization, and peer mood coupling — can run entirely on consumer mobile hardware at under 100K parameters. The Personal Arousal Function (PAF) is the interesting bit: it learns that the same track hits differently for different listeners, trained from behavioral signals like skips, volume adjustments, and favorites. If the substrate generalizes beyond music, this is a blueprint for on-device emotional intelligence in any ambient app.
Catalyst
CoreML on Apple Silicon (A16/A17) finally makes sub-100K continuous-time network inference fast enough for real-time affective feedback loops without a server round-trip. The closed-form CfC architecture, which sidesteps the ODE solver bottleneck of earlier liquid neural nets, only became practically deployable on mobile in the last 18 months.
What's New
Prior affect-aware music systems (Spotify's mood playlists, Endel) use static audio feature mappings to arousal/valence — one function fits all listeners. MeloTune replaces that with a per-listener PAF that drifts and updates live; their 46-observation deployment shows pop genre confidence converging after just 22 samples.
The Counter
The headline numbers are deceptively tidy. A trajectory MAE of 0.414 on Russell's circumplex sounds precise, but the circumplex is a two-dimensional self-report construct with notoriously noisy ground truth — users can't reliably locate their own arousal/valence state, so what exactly is the model converging to? The live deployment covers 46 behavioral observations, which is closer to a demo than a study. More structurally: Spotify has hundreds of billions of implicit feedback signals and a decade of affective playlist data. A 94K-parameter on-device model trained from a single user's skips can't compete with that prior at cold start, and the paper doesn't address cold-start degradation at all. The peer coupling story is appealing in theory, but co-listening scenarios (shared playlists, simultaneous sessions) are a small fraction of actual music consumption. Finally, the SDK and protocol (sym-swift, SYMCore) are versioned and real, but there's no evidence of third-party adoption — this could easily remain a research artifact rather than a platform.
Longs
- AAPL
- SPOT
Shorts
- Spotify (mood playlist and DJ feature moat erodes if per-listener arousal modeling commoditizes)
- Endel (their differentiation is adaptive audio; on-device PAF replicates it without their infrastructure)
Enablers (Picks & Shovels)
- Apple CoreML / Core Neural Engine
- sym-swift v0.3.78 / SYMCore v0.3.7 SDK (open protocol)
Private Watchlist
- Endel
- Suno
The Paper
MeloTune is an iPhone-deployed music agent that instantiates the Mesh Memory Protocol (MMP) and Symbolic-Vector Attention Fusion (SVAF) as a production system for affect-aware music curation with peer-to-peer mood coupling. Each device runs two closed-form continuous-time (CfC) networks: a private listener-level CfC that predicts a short-horizon affective trajectory on Russell's circumplex and drives proactive curation, and a shared mesh-runtime CfC at MMP Layer 6 that integrates Cognitive Memory Blocks (CMBs) from co-listening peers. CfC hidden states never cross the wire; only structured CMBs do. A Personal Arousal Function (PAF) replaces the standard linear mapping from audio intensity to psychological arousal with a per-listener learned adjustment, trained from behavioral signals (skip, completion, favorite, volume) and from drift between user-declared mood and machine inference. The same track receives different arousal predictions for different listeners. The model (94,552 parameters) achieves trajectory MAE 0.414, pattern accuracy 96.6%, and intent accuracy 69.4% on held-out validation. PAF evidence from a live deployment session (46 observations across 11 genres) demonstrates that the learning loop operates end-to-end, with pop reaching full confidence after 22 observations. All inference runs on-device via CoreML. To our knowledge, this is the first production deployment of MMP/SVAF on consumer mobile hardware. The accompanying SDK (sym-swift v0.3.78, SYMCore v0.3.7) enforces strict protocol conformance. Music is the case study; the substrate is the contribution.