On Machine Learning-Driven Surrogates for Sound Transmission Loss Simulations