SEAL: Speaker Error Correction using Acoustic-conditioned Large Language Models