We introduce a multimodal electromyography (EMG) inference framework trained on a collected piano performance dataset, to enable high-fidelity hand EMG estimation and demonstrate generalization and adaptability across unseen users and tasks.