Consider using a model like supertonic 3 for tts.

Its light weight and can run on web inference.