As mentioned, the text recognition part of Watson Visual Recognition cannot be trained by users.
The custom classifier learning functions cannot be used effectively for text recognition. They are designed for classifying a wide array of photographic-style images of scenes, activities, objects, animal types, etc. Text reco is a pretty different problem. In some limited situations, you could train a classifier to do text-related tasks, like to tell the difference between signs (such as Yield vs. Stop) or icons (like logos) for example but it would not be "reading" the text - it would be discriminating based on the overall appearance of the image - including color, shapes, textures, etc.