I’m calling it: The uncanny valley for voice AI is officially over.
I’ve personally been so let down by robotic-sounding AI voices that stumble over simple things like phone numbers or acronyms. It just pulls you right out of the experience. But I just saw a post from an industry pro who tested a new model that completely blew me away.
This expert had nearly given up on voice AI after spending thousands on tools that couldn’t get the basics right. They all failed the same simple tests, sounding unnatural and robotic. Then, the creator tested Cartesia’s new Sonic-3 model, and the difference was night and day.
Here’s what this talented creator discovered:
- 📌 It nails the basics. The original poster shared examples where Sonic-3 reads phone numbers, acronyms (like “FBI”), and complex addresses naturally, just like a person would. No more robotic “Five. Five. Five.” or spelling out letters.
- 💡 It handles complex sentences. The mind behind it threw rapid-fire, complex sentences at the model, mixing names, numbers, and acronyms, and it didn’t stumble. The voice remained smooth and natural without any weird pauses or hesitations.
- ✅ It has next-level dynamic abilities. This is the really wild part. The person who shared it asked the AI to laugh, and it laughed naturally. It could adjust its speaking speed in real-time and even translate a sentence into another language while keeping the same voice and accent.
This isn’t just another minor update; it’s what enterprise voice AI should have been from the start. Voice agents that actually sound human will completely transform customer experience.
The original post has the full breakdown of the tests. You have to check it out to understand how good this is!