Everyone's first instinct for making an AI sound human is the same: sprinkle in some "ums" and "uhs" and a few "mhmm"s. We tried it. It sounded worse.
Through a text-to-speech voice, scripted filler words land in the wrong spots — an "um" arrives too crisp, too on-beat, and the whole thing tips from "polished robot" into "uncanny robot." A tester noticed immediately, and so did the client.
The real reason an agent sounds robotic
It isn't the lack of filler words. It's the too-perfect, even pacing of a synthetic voice — every syllable equally weighted, no rush, no drift. That flawless rhythm is the tell.
What actually helps
A casual, conversational voice instead of a polished "narrator/announcer" one. Voice character matters more than any single setting.
A little more looseness and variation in the delivery, so the rhythm isn't machine-even.
Writing the way people talk — contractions, short sentences, natural pauses — instead of tidy paragraphs.
And one honest thing we tell every client: there's a ceiling. Real-time voice trades a sliver of realism for speed, so on a dead-quiet line a sharp listener may still clock it. That's fine — the goal is a warm, easy call, not a trick. (We also never have an agent deny being AI if asked. Sounding human and pretending to be human are two very different things.)
The owner takeaway
Judge a voice demo by one question: does it sound like a relaxed person on a normal call? Not by whether it remembered to say "um."
Want to hear what your line could sound like? Grab a free workflow audit.