how can I get my project reviewed by an IRB?
You pay them. University IRBs may or may not handle cases from outside the university, but there are commercial IRBs that do this all the time.
It is likely that your project would qualify for a "minimal risk" exemption (e.g., if you are having participants complete simple tasks and recording their responses and reaction times, then the odds of a participant being harmed are negligible). But, even this would be take months and cost at least $1,000.
The possible complication is that you would presumably be recording diagnoses (e.g., which participants are dyslexic) and publishing data based on this dataset. I do not know if this would be considered a "medical record," but I suspect it might be, in which case you might actually need a full IRB protocol. This would probably be a no-go: I seriously doubt any IRB would approve a study in which the principal investigator is a high school student. (And even if they did, this would be very expensive and time consuming).
one of my teachers told me that I need IRB approval.
This may or may not technically be true. For sure federally-funded research requires an IRB (either a full protocol or a formal exemption), and if you want to publish this in a "real" science journal, they will require it too. If this is for a science competition, they may have a policy requiring one (though they probably just say "at your own risk"). If none of the above apply, you probably don't actually need IRB approval.
That said, the above discussion should hopefully make it clear that there are lots of legal landmines here. While the odds of you harming someone and getting sued are very small, it could really mess up your life if it happened. Your teacher advisor is probably (rightly) concerned about this (I suspect "you'd need IRB approval for that" is an elegant way of saying "no freaking way!").
I want to improve on existing machine learning models for detecting dyslexia. I want to collect data from both dyslexic and non-dyslexic participants.
Bear in mind that machine learning models are data-hungry, and collecting data is tedious. Even getting data from 50 participants would likely require a heroic amount of work on your part, and it's unlikely you could do much machine learning based on such a small dataset.