Patients with neuromuscular diseases such as amyotrophic lateral sclerosis can have difficulty communicating because neural signals cannot reach effector muscles. Recent advances in brain-computer interfaces have allowed these patients to communicate by converting neurological signals into computer commands. One common brain-computer interface is the P300 speller, a system that allows these patients to spell out text. Because of the electroencephalogram (EEG) signal variability between patients, it is hard to create a classifier applicable to all patients. Therefore, current methods use an arduous training step personalized for each patient. There have been previous attempts to create a general classifier that works for all subjects, but these attempts have generally resulted in poor accuracies that were insufficient for practical use. This paper presents a novel cross-subject approach for the P300 speller. It uses a language model which adjusts the probabilities of each character based on context to improve classifier performance. Additionally, dynamic stopping allows the system to continually obtain EEG signal from the patient until the system is confident in its character selection. By using these two approaches, we can maintain reasonable selection accuracy, allowing subjects to use the system without an individualized training step.