Edge-aware Emotion-adaptive Music Playback: A Prototype Reference Architecture with MobileNetV3 Facial Mood Sensing and Hybrid Context-aware Ranking

Abstract

Music recommendation systems typically model long-term taste more effectively than moment-to-moment affective context. This paper presents a prototype reference architecture for emotion-adaptive music playback that integrates three artificial-intelligence components: MobileNetV3-based facial emotion recognition (FER) for low-compute mood sensing, CNN/Transformer-based music emotion recognition (MER) for song affect representation, and a hybrid context-aware ranker that combines collaborative preference, audio/mood features, and current affect. Frame-level FER posteriors are temporally stabilised with exponential moving average smoothing before they are mapped to discrete emotions or valence-arousal coordinates. Songs are designed to be indexed offline from log-mel spectrograms and complementary audio descriptors, while the ranking layer operates at playback time. The architecture separates local affect inference from server-side personalisation so that raw facial frames need not be retained. The contribution of the present study is architectural and methodological rather than a population-level empirical claim: no human-participant dataset, real behavioural interaction dataset, or named edge-hardware benchmark was used to generate the numerical examples in this manuscript. Accordingly, the metric values shown in the illustrative benchmark tables and figures are synthetic engineering examples used only to demonstrate the proposed evaluation, ranking, and ablation workflow; they are not measurements of user behaviour or device performance. The paper therefore contributes a mathematical formulation, implementation-ready module and database design, privacy requirements, reproducible empirical validation protocol, and a prespecified statistical analysis plan for subsequent FER, MER, hardware, ranking, and user studies.