Afaik its only gonna be used for stuff like screen reading, which if you've ever tried an open source speech synthesis model, you'd know even an old lightweight LLM model is better than it.
I'd also argue that if you actually care about local llms, you can just set up ollama and use that.