Is it worth building a handwritten text recognition model for a mobile app (need to finish this project by next month)? Must be at least 75% accurate.

I’m building a simple Android mobile app (react native front end) capable of recognizing handwritten text as part of my school project.

Basically, the concept is that there are 15 phrases that will be written by a child in a paper and the app must properly recognize it at least 75% of the time. The phrases are simple like: “hello”, “good morning”, “how are you?” and so on.

It’s not for general handwritten text recognition purposes but rather there are only 15 specific phrases that needs to be recognized. The ones writing those phrases are children aged 10-14.

What’s the best option here cost and performance-wise? Should I just avail API such as Google Vision or train a model, such as the one with PaddleOCR or TrOCR, and deploy it locally? If the latter, what platform can I train on and what dataset can I use specifically for my case?

I tried the Google Vision and Microsoft APIs for text recognition and they did perform well but it costs around 2$ for 1000 requests. I also tried Tesseract and PaddleOCR inference models but they didn’t perform well.