To help prevent malicious use, consider presenting the user with specific (randomly-generated) text to read aloud, and check (with speech to text) that they act

To help prevent malicious use, consider presenting the user with specific (randomly-generated) text to read aloud, and check (with speech to text) that they actually read that, instead of allowing them to say whatever they want. That will help ensure that this is only being used by the person visiting the web page. (That will only help with the hosted version, of course, not if you make the model code/weights available. I didn't generate this idea myself but also can't remember where I saw it. I think it was from someone offering a similar service.)