講演情報
[OS1-6-02]Brainwave Entrainment via Generative Audio Models
Neupane Oshika2、Maharjan Simran2、K.C Yugesh2、伊藤 篤1、平松 裕子1、*Khansakar Anila2 (1. 中央大学、2. Tribhuvan University)
キーワード：
EEG、Meditation、Brainwave Entrainment、MusicGen、Reinforcement Learning
We present an EEG-informed generative AI system that fine-tunes a pretrained music generation model to produce audio associated with increased alpha-band activity (8–13 Hz), linked to relaxed wakefulness. Unlike conventional approaches that use static audio, our method combines consumer EEG data with preference-based reward model- ing and fine-tunes MUSICGEN via reinforcement learn- ing (RL). EEG data from 49 participants across 43 tracks provide a reward signal based on normalized changes in alpha-band power. A reward model trained via pairwise preference learning predicts the alpha-inducing potential of generated audio and is integrated into a REINFORCE- based framework with LoRA to guide the generation pro- cess. Results show that the reward model loss decreases from 0.90 to 0.34, and RL rewards improve over 7,000 steps. These results demonstrate that physiological signals, such as EEG, can effectively guide generative audio models, thereby enabling a data-driven link between brain activity and music generation.