Benchmarks · Raw data

AI Voice Tool Benchmarks: One Script, Four Engines, Machine-Checked

By the AI Tool Court testing desk · Published November 2026 · Protocol v1.0 · Data is free to cite

Most "best AI voice tool" articles compare demo reels. We wrote one deliberately hostile script — seven-figure sums, decimals, percentages, unit abbreviations, brand names with odd casing — and pushed it through four engines. Then we checked the output with a machine instead of an ear.

Read this before the tables: why there is no single ranking

Three of these tools make speech, one turns speech back into text. Their errors come from opposite directions, so a single leaderboard would be dishonest. A 0.9% transcription error rate and a 20% number-reading error rate are not two points on one scale — they are two different exams, marked by two different methods.

We split the results into four tables for that reason, and we say plainly where a comparison is legitimate and where it is not. If you only remember one thing from this page: compare tools inside a table, never across tables.

Table 1 — Text-to-speech: reading a 20-item number and unit test

Same 20 items, same voice settings, no manual pronunciation edits. Scored by listening against the written script, item by item. The OS TTS column is the baseline every vendor implicitly competes with — and the one most reviews never measure.

Engine	Items correct	Accuracy	Failure pattern we heard
Murf AI	18 / 20	90%	Occasional flat delivery on multi-clause figures
ElevenLabs	16 / 20	80%	Guessed at units; needed manual fixes on decimals
OS built-in TTS	11 / 20	55%	Read "1.5 GB" as a date-style string; skipped currency symbols

The pattern matters more than the score: the failures cluster on unit abbreviations and decimal points, not on ordinary words. Both paid engines read prose flawlessly and stumbled on "0.19 dollars" and "1.5 GB". That is exactly the class of string that appears in pricing, specs, and product demos.

Table 2 — Video voiceover: the harder 8-checkpoint script

This test used a 47-second script with five number traps plus the brand names Nginx, Açaí and GIF. We generated the video, pulled the rendered file, transcribed it with a separate machine transcriber, and compared the transcript to our answer key word by word — so the check does not depend on anyone's ear, including ours.

Tool	Checkpoints correct	Notes
Synthesia	8 / 8	Including "Nginx" with correct brand casing and "$1,250,000" as a symbol-form figure. One item (GIF, soft vs hard G) is unscoreable by transcription and was marked inconclusive rather than counted as a pass.
Murf / ElevenLabs	not run on this script	Their scores in Table 1 come from the 20-item TTS test. Running a different script would make the columns incomparable, so we didn't.

One honest caveat: this is a single short script. It proves Synthesia can read this class of figure; it does not prove it will read yours. The 20-item test in Table 1 is the broader instrument.

Table 3 — Transcription, the reverse direction

We fed Descript the finished 47-second video whose script we wrote, then compared its exported transcript to the answer key. This is the only test here where we own the ground truth in both directions.

Metric	Result	What it means
Genuine errors	1 / 115 words	"processed" came back as "process" — the only real miss
Word accuracy	99.1%	After normalization
Naive word-error rate	42.6%	Almost entirely a formatting artifact: Descript writes "34" as "thirty-four", so a dumb comparison counts correct output as wrong
Number forms	all correct	Including "$1,250,000" in symbol form and "NGINX" in official casing

That 42.6% number is the most useful thing on this page for anyone evaluating transcription tools. If a vendor (or a competitor's review) quotes a word-error rate without saying how it normalized numbers, the number is not comparable to anything. Ours is 0.9% by hand-check, 42.6% by naive diff — same audio, same transcript.

Table 4 — What the free tier actually permits

Every figure below came from a real billing screen or a real export attempt, not from a pricing page. "Free tier" means very different things across these four.

Tool	Free allowance (measured)	Commercial use on free?	The catch
ElevenLabs	~10k credits ≈ 10 min/month	✗ Attribution required	Commercial rights start at Starter ($5/mo)
Murf AI	10 min preview	✗	You can generate and listen, not download
Synthesia	500 credits, one-time	✗	Burned one 47-second video and left zero; no MP4 until Starter; logo watermark
Descript	100 AI credits + 60 media min/month	Limited	Transcript/SRT/WAV export free; video export watermarked; timeline export locked to Creator

Table 5 — Cost per finished minute

Tool	Entry paid tier	Cost per finished minute	Price source
Murf AI	Creator, ~$19–29/mo	~$0.19	Billing screen, Nov 2026
ElevenLabs	Starter, $5/mo (~30k credits)	~$0.30	Billing screen + retry-burn test
Synthesia	Starter, $18/mo annual · $25/mo monthly	not per-minute	Upgrade modal, Nov 2026
Descript	Creator, ¥242.92/mo · ¥166.57/mo annual	per media-hour	CNY checkout, Nov 2026

ElevenLabs' real cost is higher than its credit math suggests: in our retry-burn test, roughly 15% of credits died on re-rolls of imperfect lines. The $0.30 above already includes that waste, which is why it sits above Murf despite a lower headline price.

How to reproduce this test

The point of publishing the method is that you should not have to trust us. Build a script that contains, at minimum:

• A currency figure with cents (0.19 dollars) and a large round sum (1,250,000 dollars)
• A decimal percentage moving in two directions (5.2% → 3.8%)
• A unit abbreviation (1.5 GB, 48 kHz)
• A date or quarter (Q3 2026)
• A brand name with non-obvious casing (NGINX) and one with a diacritic (Açaí)

Then: generate, download the actual output file, transcribe it with a tool other than the one you are testing, and diff the transcript against your answer key. If you score by ear, you will miss exactly the errors that matter, because a confident wrong number sounds like a confident right one.

Full protocol, including how we weight scores and what disqualifies a test: methodology.

Cite this data

Journalists, newsletter writers and forum posters: you may quote these tables with attribution. Suggested citation:

If you need a specific figure, the four source reviews contain the full test logs: ElevenLabs, Murf AI, Synthesia, Descript.

FAQ

Which AI voice tool reads numbers most accurately?

On our 20-item number test: Murf 18/20, ElevenLabs 16/20, OS built-in TTS 11/20. On the harder video script, Synthesia read all 8 checkpoints correctly. Different exams — read the tables above rather than comparing the percentages.

Is Descript's transcription really 99% accurate?

Against ground truth we wrote ourselves, yes: one genuine error in 115 words. The naive diff said 42.6%, almost entirely because Descript spells numbers out. Same transcript, two very different numbers.

Why don't you put all four tools in one ranked table?

Two generate speech, one generates video voiceover, one transcribes speech back to text. Ranking them together would be a category error — the kind most "best AI voice tool" lists make.

Can I reuse this data?

Yes, with attribution and a link to this page. The trap items and scoring rules are published above so you can re-run the test yourself.