A large collection of research literature has identified the privacy risks of keystroke inference attacks that use statistical models to extract content typed onto a keyboard. Yet existing attacks cannot operate in realistic settings, and rely on strong assumptions of labeled training data, knowledge of keyboard layout, carefully placed sensors or data from other side-channels. This paper describes experiences developing and evaluating a general, video-based keystroke inference attack that operates in common public settings using a single commodity camera phone, with no pretraining, no keyboard knowledge, no local sensors, and no side-channels. We show that using a self-supervised approach, noisy finger tracking data from a video can be processed, labeled and filtered to train DNN keystroke inference models that operate accurately on the same video. Using IRB approved user studies, we validate attack efficacy across a variety of environments, keyboards, and content, and users with different typing behaviors and abilities. Our project website is here.