If you are building this out, I recommend starting with the first, as it doesn't require extra hardware permissions. You can use an LLM (like GPT-4o) to pre-scan the text for "Perplexity" scores and map the speed changes accordingly. If you'd like to explore more, let me know: Is this for a mobile app or a browser extension ?
Speechify’s free tier is still one of the best in the market. It provides high-quality standard voices that are perfectly functional for students and professionals.
Please provide more context, and I'll do my best to assist you.