Voice is where a lot of the practical AI product work is heading, and the difficulty is not the models, it is the plumbing. You need transcription, then a model, then synthesis, and you need all of it fast enough that a human does not notice the machinery. Cartesia has built its documentation around exactly that problem, which makes it more useful than a typical model provider's reference. The stack as documented is Sonic for text-to-speech, positioned around roughly ninety milliseconds of latency and emotional expressiveness, Ink for speech-to-text with turn detection built in, and Managed Agents as a platform combining transcription, a model of your choosing and synthesis with turn-taking and telephony handled for you.
Turn detection inside the transcription model is the design decision worth understanding, because knowing when a human has finished speaking is the problem that makes voice agents feel broken. Bolting a separate voice activity detector on top produces the awkward pauses and interruptions everybody has experienced on an automated call. Building it into the model is the right architecture and the docs explain why. The playground is the best feature of the whole documentation experience.
Voice quality is subjective and unarguable in a way that text quality is not: you either believe the voice or you do not, and no benchmark will settle it. Being able to hear the output in thirty seconds, before creating an account or writing a line of integration code, is exactly right and more vendors should do it. Stated latency figures deserve credit too. Ninety milliseconds is a specific claim you can go and verify, which is more useful than the industry-standard adjectives.
The Managed Agents path is a sensible product for teams who want a voice agent rather than a voice infrastructure project. Telephony alone is a swamp, and handing it over is defensible. The trade is the usual one: the abstraction is comfortable until you need to do something it did not anticipate, and then you are learning the underlying system anyway, under time pressure, with a product already shipped. The documentation does not flag that trade, and a paragraph on when to drop to the lower-level APIs would serve readers well.
Voice cloning is where I would push back hardest. It is documented as a capability alongside the others, technically and matter-of-factly, and it is not like the other capabilities. Cloning a voice raises consent questions, misuse questions and, in a growing number of jurisdictions, legal ones. Fraud using cloned voices is not hypothetical, it is a live and growing category of crime.
A documentation set that explains how to clone a voice at the same register it explains how to set a sample rate is making an editorial choice, and I think it is the wrong one. This is not unique to Cartesia and it is not a reason to avoid the product, but the industry norm here is worse than it should be and reviews should say so. Pricing, as with nearly every provider in this batch, is absent from the documentation. The playground gets you an API key and the costs live elsewhere.
The other gap is failure handling: what a well-built voice agent does when transcription is uncertain, when a caller interrupts mid-sentence, when the network degrades. That is where voice products are actually won and lost and the documentation is largely silent on it. Three point seven. Clear, well-structured documentation for a genuinely good real-time voice stack, with a playground that makes honest evaluation easy, marked down for treating voice cloning as a routine feature and for skipping the failure cases that determine whether a voice agent is usable.