Can you pick the paid voice blind?
Creative AI Tests, episode three: Piper and Kokoro running free on my own machine against the paid voice I already use. Four voices, one script, no labels until the reveal.

Four voices read the same twenty seconds of text, and only one of them costs me money. That was the whole test this week, and I'll be honest: before I ran it, I wasn't sure I'd pass it myself.
Creative AI Tests is my Sunday series. I take one tool, give it a real job and report back on what happened. Episode three was text to speech, free against paid, and the question fits on a sticky note. If you couldn't see the labels, would you know which voice I pay for?
The setup
On the free side were Piper and Kokoro. Both run locally on my own machine, so each line costs nothing beyond the power bill and a bit of fan noise. From Piper I used two voices, an American one called lessac and a British one called alba. From Kokoro I used a voice called am_adam. The paid side was a voice tool I already pay for. I'm not naming it, partly because the brand isn't the point and partly because "the paid one" keeps the guessing honest.
I trimmed the script to about twenty seconds. Short enough to keep the comparison fair, and short enough that the paid read stays cheap. Every voice read these exact words:
"Kia ora. This is a short voice test. I'm reading the same lines so you can compare how natural each voice feels. Listen for the pauses, the consonants, and whether it still sounds human when I pick up the pace near the end."
That last sentence is a small trap. Plenty of voices sound fine reading slowly. Speed is where the seams show.
Each clip was edited and level-matched so nobody won just by being louder. Then they went into the video as Voice A, B, C and D, with nothing on screen to give the game away.
The reveal
Voice A was Piper. Voice B was Kokoro. Voice C was Piper's British voice, and Voice D was the paid one.
If you picked it, I'd bet the clue came from the pauses and the pace rather than the vocabulary.
Where the free voices give themselves away
Soft consonants blur together. Not badly, just enough that the edges of words go fuzzy, like someone talking with a cup of tea held near their mouth.
Longer sentences settle into a flat, even rhythm. The voice starts with some life, finds a cruising speed and stays there, which is lovely in a car and less lovely in a narrator.
Emotion is thin too. If a line has a smile in it, the smile doesn't quite land. And when two free voices sit back to back, you start to hear the same cadence repeat, as if they went to the same elocution teacher and both paid attention.
Names and place names are the other usual weak spot. They can come out stiff even when the words are right. I'll put my hand up here, though: names weren't tested in this sample, so treat that one as a warning rather than a result.
Mr. Mummu runs his own test
Mr. Mummu, who turns up in these posts now and then, heard about all this and decided to run a blind test of his own, on doorbells. He recorded four of them, labelled the files in his notebook marked "Urgent", then shuffled them so thoroughly he forgot which was which. The notebook page now says "B is the good one?" with the question mark gone over three times.
He installed B anyway. The first evening someone pressed it, he didn't get up. It sounded like the oven timer, so he waited for it to finish.
The courier left a card.
And that's what a blind test teaches you. Nobody needs to know which product you bought. They only need to notice that something sounded flat, and then they stop listening.
Where free is fine
None of this makes the free voices useless. I'd happily use them for drafts, internal explainers, scratch reads while I'm shaping a cut, and rough tracks to get the timing right. Anything that never goes out as the final public voice is fair game, and at zero cost per line you can be as wasteful as you like.
For a LinkedIn clip or a YouTube episode with my name on it, I still want the paid voice. Viewers forgive a lot. A wobbly cut, a so-so thumbnail, a slightly long intro. But they notice a flat read.
So my takeaway is short. Use free text to speech for drafts, and keep the paid voice for anything public.
The honest limits
This was one short script, read once by each voice and judged by my ears. A different script, or a sample full of names, could shift things. I'd call it a useful first look rather than a verdict.
You can watch the full test, with all four voices and the reveal, in about four minutes on my YouTube channel, Gautam Atmakuri, in the Creative AI Tests playlist. I'll put the link in the comments.
Next up is free speech to text on a Kiwi accent, once I've got a proper sample. Place names and all.
Until then, listen to the end of the sentence. That's where the money is.
About the Creator
Gautam Atmakuri
I once spent seven days blindfolded, silent and without food to see what my senses were hiding. I write about that, my baby boy, barefoot running, Tirupati to Auckland life, and the business, marketing and tech I build at AumnI digital.
Enjoyed the story? Support the Creator.
Subscribe for free to receive all their stories in your feed. You could also become a paid subscriber, letting them know you appreciate their work.
Comments
There are no comments for this story
Be the first to respond and start the conversation.