Do AI receptionists sound human?

Text-to-speech has improved enormously. The flat, clipped delivery people remember from phone menus is gone, and pauses and emphasis land in roughly the right places.

What gives it away is rarely the voice itself. It is the turn-taking: a beat too long before the reply, or an answer that starts before you finished the sentence.

Callers also notice consistency. A human receptionist sounds tired at five o'clock and cheerful at nine. An assistant sounds the same at both, which is pleasant and slightly uncanny.

Maine's Title 10 §1500-DD took effect on October 1 2025. It bars using an AI in trade or commerce so as to mislead a reasonable consumer into thinking they are dealing with a human, unless the consumer is clearly and conspicuously notified. The definition covers aural interactions, so an inbound phone assistant is squarely inside it.

That makes any marketing claim of the form "your callers will never know" an odd thing to advertise. A vendor selling you on not being detected is selling you the behavior a statute names.

Utah goes further for some businesses. Under its AI Policy Act, regulated occupations such as medicine, law, accounting, nursing and psychology must disclose verbally at the start of the interaction.

There is a practical argument too. A caller who works out mid-sentence that they were misled gets annoyed at you, not at the software.

Say you manage a veterinary clinic and you want to judge the voice properly before it meets a worried pet owner. Don't listen to a sample clip. Place four calls from a real phone instead, because a phone line carries a much narrower band of sound than the speakers on your laptop, and a voice that sounds rich in a browser can sound thinner over the network.

Call one is the easy booking: a vaccination next week. On call two, interrupt it halfway through a sentence and see whether it stops talking or plows on. For call three, ring from a car on speaker with the window down. On call four, give it a hard name, something like Siobhan Nguyen at 14 Schuylkill Court, and listen to how it says the name back when it uses it.

You're listening for the gap before each reply and for what happens when two people talk at once. Those two moments carry most of the impression, and no adjective on a sales page can tell you how they'll feel.

Your script moves the needle more than the voice model does. A sentence of 35 words with two sub-clauses sounds read aloud no matter who, or what, is reading it. Cut the greeting to something you'd say yourself across a counter and the same voice sounds far more natural.

Local words are the other weak spot. Street names, staff surnames, a suburb with an odd pronunciation: any synthetic voice can trip on these, so put the awkward ones into a test call early. If your town is pronounced differently from how it's spelled, find out in week one rather than from a customer.

None of this changes the disclosure question. In Maine, and for state-licensed professions in Utah, the caller has to be told they're speaking with an AI however natural the voice is, so treat naturalness as a comfort feature and never as cover.

GreetKeeper can announce itself as an assistant in the opening line, and you write that line. Something plain works best: the practice name, then that you are speaking with its assistant, then a question.

You can hear it before you buy. A demo runs your own scenario, which tells you more than any adjective we could put on this page.

The transcript shows you where the conversation stalled, so the fix is usually a wording change rather than a voice change.