How to Compare AI Narrators on the Same Chapter: A Free Bake-Off You Can Run

You can compare AI narrators for free by running one short passage of the same chapter through each tool's free tier, with identical settings, and scoring the results blind. NarratedIndie has not published audio samples yet, so there is nothing to play on this page; what follows is a method you can run yourself.

Which free tiers allow a bake-off, and their limits

The free allowances below, as listed on 2026-11-08, are enough for a short passage, and the output is non-commercial where the vendor says so. Do not use free-tier output where the vendor says it is non-commercial.

Murf's free plan could not be confirmed, because its pricing page returned no readable content on 2026-11-08, and third-party figures conflict. Check it live. See the Murf review.

Free tiers may differ from paid output, such as ElevenLabs' 128 kbps free audio against higher listed quality on paid plans, so what you hear on a free plan may not match what you would get when paying. Treat a free bake-off as a shortlist step.

How many runs fit in each allowance

A 400-word passage leaves room for very few retakes on most free tiers, so plan your runs before you spend them. These are our estimates, using 400 words × an assumed 6 characters per word = 2,400 characters, and 400 ÷ 9,300 words per finished hour = 2.6 minutes, or about 155 seconds.

ElevenLabs' Studio docs say the Regenerate button can be used twice without credits if the voice and text do not change. Its pricing page says only that a limited number of free regenerations may be available, so watch your balance.

Keep the field to three or four voices. More than that is hard to hold in your head, and each one adds listening time.

Before you listen: remove the tools you could not use

Rule out any tool whose file cannot reach your store, because a voice you like is no use if the route refuses it. As read on 2026-11-08, Voices by INaudio accepts only unmodified LPF packages from Google Play Books, ElevenLabs and Spoken Press, and Spotify's digital voice page names Google Play Books and ElevenLabs "and more". A tool a store does not name is not named, which is different from refused. See Voices by INaudio, and write the route beside each candidate before you start.

Step 1: choose the passage

Pick a public-domain chapter or one of your own, with dialogue, names and hard words in it. Public-domain status depends on the country, so check it for yours. Cut the passage to about 350 to 400 words. At ACX's old 2013 figure of about 9,300 words per finished hour, that is roughly 2.5 minutes, our estimate, which fits the WellSaid 3-minute trial with a little room. Actual length varies by voice and pacing.

Before you start, mark the passage for known trouble:

• character names and invented words;
• numbers, dates, abbreviations and units;
• words spelled the same but said differently, depending on context;
• a long sentence, a short punchy one, and a list;
• dialogue with two speakers and a speaker tag;
• one emotional beat and one plain descriptive stretch.

Do not paste someone else's copyrighted chapter into a tool whose terms you have not read.

A practice passage you can paste

If you want a stress test to sit alongside your own text, use the passage below. We wrote it for this page and it contains each trouble spot above. It is about 170 words, so add about 200 words of your own book to reach the target length. It is a script only; we have not generated audio from it.

Dr. Aoife Brannigan closed the ledger at 9:05 on the morning of 3/4/1984, and the lead weight she had used to hold its pages rolled off the desk. The St. Anne's Rd. clinic owed $1,250, which was 40% more than it had earned since the flood, and the flood was the reason she had stopped sleeping, stopped answering letters from the Kestrelmoor Trust, and started carrying the ledger in her coat.

"You read it again," Tomasz said from the door.

"I read it every day." She did not look up. "Close the door, please. You're letting the cold in."

He came close enough to see the figures. "Is that a tear in the page?"

"It's a tear," she said, and her voice broke on the word, and neither of them pretended it hadn't.

The back room held three things: a kettle, a folding cot, and forty-one boxes of unsorted files. The window faced east. In winter the light reached the far wall by ten and was gone by noon.

What to mark on your copy:

• Abbreviations: "Dr." and "St." and "Rd."
• Numbers: 9:05, 3/4/1984, $1,250, 40%
• Names: Aoife Brannigan, Kestrelmoor, Tomasz, and decide how each should sound before you listen
• Same spelling, different sound: lead, read (twice, past and present), close (twice), tear (twice)
• Long sentence: the second one
• Emotional beat: "It's a tear"
• List and plain description: the final paragraph

To check how a voice holds up across a book, add a second short passage from a chapter far from the first and generate it in a separate run.

Step 2: keep the settings identical

Give every tool the same text file, with the same pronunciation fixes applied or none at all. Use the same speaking speed setting if the tool offers one, otherwise leave defaults and record them. Pick one voice per tool, chosen before you listen to the others, and do not change it. Record for each run:

• the tool, plan, date and voice name;
• the model or version if the tool shows one;
• the credits, minutes or characters used;
• the number of times you regenerated it;
• any edit you made to the text and why.

If you retake a sentence, log it. Retakes are part of the cost, as the cost calculator shows.

Step 3: blind the files

Have someone else rename the files A, B, C and so on, and keep the key away from you until you finish scoring. Export at the same format where possible and match loudness so that louder is not heard as better. Listen on the headphones or speakers your readers are most likely to use. Listen twice, on different days if you can, and note what changed.

Free tiers can break the blinding, and it helps to know where. Speechify's free plan does not include MP3 download, so a helper would have to play it to you from the app. Spotify's voice picker plays samples inside its own interface. Narration Box's free output is watermarked and 8 kHz, so you will recognize it at once. Score those candidates knowing the comparison is weaker, or leave them out of the blind round.

If you are working alone, have a spreadsheet's random function assign the letters, and save the key in a file you do not open until scoring is done. A second or third listener is worth more than a second listen by you, so ask a reader or two if you can.

Step 4: what to listen for

Score what affects an audiobook listener over many hours, not a single clip:

1. Pronunciation. Count errors on your marked words and on any others you hear.
2. Pacing. Check pauses at commas, paragraph breaks and dialogue changes.
3. Dialogue. Can you tell who is speaking without the tag?
4. Consistency. Does the voice stay the same from the first sentence to the last?
5. Artifacts. Listen for clicks, glitches, repeated or skipped words.
6. Fit. Does the delivery suit the genre and the emotional beat?

Step 5: score and keep a log

Give each criterion 1 to 5 and also count hard errors per 400 words. Add the minutes you would need to fix the problems, because edit time is a cost. A table like this is a template for you to fill in with your own results:

Add a column for the store routes you need. The three-way comparison lists them, and best AI audiobook generators explains how routes affect the choice.

Turn the scores into one number

A weighted total stops one striking voice from winning on charm alone. Set your weights before you listen, so they add to 100. The weights below are a suggested starting point, not a standard. For nonfiction with no dialogue, move the dialogue weight to pronunciation.

Weighted score = the sum of (score × weight) ÷ 5. The numbers below are invented to show the arithmetic. Voice A and Voice B are not real products.

A one-point gap on a single listener's scores is a tie. Treat anything within about five points as a tie, which is a rule of thumb of ours, and let the next step decide it.

Round two: can you fix what went wrong?

Once you have unblinded the results, take the top two voices and try to correct their three worst errors with each tool's own controls. A voice with more errors may still be the better choice if its errors are easy to correct. The controls differ widely; see pronunciation and pacing fixes for what each one documents. Note whether each fix worked, how many attempts it took, and what it cost in credits.

Then break any remaining tie on things that are not sound: which route the file can reach, and what the book would cost to finish on a paid plan.

Limits of this method

A short passage will not show how a voice holds up over a whole book, and one listener's scores are one listener's scores. This method also does not tell you whether a listener would notice AI narration at all; for that, see what published studies measured in Can Listeners Tell AI Narration?. When you have a shortlist, read the reviews for ElevenLabs and Speechify Studio, and return to the AI audiobooks hub for the rest of the workflow.

Keep bake-off files private. Free-tier audio is non-commercial on several of these plans, so do not post it to promote your book or to show readers your comparison.

Last verified 2026-11-08. Rules and prices change: check the linked primary page before you rely on this.