Almost every text-to-speech product is an API with a UI on top. You paste text, it goes to a server, audio comes back, and you are billed per character. AudioColor works the other way around: the model runs on your machine.
That choice has a real cost, and it is worth being straight about it.
The cost: the first download
Voice models are large. The first time you use one, you wait while it downloads, and it takes real disk space. There is no way around that — it is the price of having the model locally at all.
What you get for it
No per-character billing. Once the model is on disk, generating a 90,000-word audiobook costs exactly as much as generating one sentence: nothing. This is the difference that matters most for long-form work, and it is the reason the pricing page has no usage tiers on it.
Your script does not leave the machine. For a lot of people this is a compliance question rather than a philosophical one — an unreleased script, a client's internal training material, a manuscript under NDA. Local synthesis means there is no upload to explain to anyone.
It works with no network. On a plane, on bad hotel wifi, in a room with no signal. The app does not degrade when the connection does.
No rate limits and no queue but your own. The only thing deciding how fast your renders go is your hardware, and the concurrency limit you set yourself.
The one exception
Automatic casting — the feature that reads a novel and proposes which character should get which voice — is the exception, and it is opt-in for that reason. It calls a language model, and language models good enough to do that job well do not fit the same local constraints yet. So it needs your own API key for Anthropic, OpenAI or DeepSeek, and when it runs, the relevant text goes to whichever of those you configured.
If you never turn it on, nothing about your documents ever touches the network. That boundary is deliberate, and it is drawn where it is so you can decide which side of it you want to be on.