Does training the model with one letter images helps ?

[Alkabann]

EasyOCR Github provide a traning test of 1000 images to train a custom model. I saw that they use one or two words for each image. Is it a bad idea to give the model letters to train if, at the ends, we want to predict words ?
If I have the image with "there is a game", should I only create the images "there", "is" and "game" ?
Actually I'm creating the images:
"there is a game", "there", "is", "a", "game", "t", "h", "e", "r", "e", "i", "s", "a", "g", "a", "m", "e"

I did some test with and without the letters, I passed from 76 to 78% of accuracy but this is kinda biased because I have less than 1500 images to train so I'm hopping to find someone that already asked himself this question and could help me understand.

I feel that it could help by understanding the concept of letters in the words but maybe giving the letter the same size of a word in the dataset makes it useless or some little tips like this.