Research

Transcription accuracy compared

6 services publish a figure for accuracy, and the word up to appears in most of them. We read every claim on the service's own page on 25 October 2026, and we publish no percentage of our own, because our pages do not state one.

Updated 2026-10-25 · Author: Mark Burmistrov · 9 minutes

What each service publishes

There is a number on the front page of most of these services, and it is the first thing a buyer reads. It is also the number a buyer has the least ground to check, because the test behind it is not published. So this page does one narrow thing: it goes to the page where the number stands and copies the sentence around it, including the conditions the service attaches. Where the number is a claim about people rather than about an engine, we say that too. Every row carries the address of the page and the day we opened it, and we opened all of them on 25 October 2026.

The accuracy figure published by each service
ServiceThe published figureWhat the service says it depends onWhere we read it
VEED up to 99.9% on automatic transcription the same page says results vary with audio quality and speech clarity, so a few words may need correcting from time to time (veed.io, transcription, checked ) veed.io, transcription, checked
TurboScribe 99.8% the page says the accuracy varies by language and lists the ones where the result is best, Russian among them (turboscribe.ai, pricing, checked ) turboscribe.ai, pricing, checked
Sonix up to 99% for clear audio recordings the page names the four things the result depends on: audio quality, background noise, speaker clarity and vocabulary (sonix.ai, automated transcription, checked ) sonix.ai, automated transcription, checked
Rev 99%+ accurate, guaranteed in 12 hours or less this is a promise about a human service, not about an automatic engine, and the figure stands in the link to that service on the home page (rev.com, home page, checked ) rev.com, home page, checked
HappyScribe the navigation claims 99% accurate, and a chart on the home page puts English at 93% against 88% for other tools the chart says it is up to 15 points more accurate than other tools, language by language (happyscribe.com, home page, checked ) happyscribe.com, home page, checked
Descript up to 95%, with clean transcripts that need minimal manual fixes the page pairs the figure with a claim of industry-leading precision across 25 languages (descript.com, transcription, checked ) descript.com, transcription, checked
Transcrib Si the source does not report a percentage our pages describe what a segment carries, the time it starts and ends and the speaker it belongs to, and stop there (the general guide, checked ) pricing, checked

Three services that promise accuracy without a number

The other half of the picture is what a service says when it says nothing measurable. A phrase like industry leader in transcription accuracy is not a figure, and it cannot be checked against anything. We put those pages in a second table rather than dropping them, because a reader who lands on a competitor's home page should know that the promise there is unnumbered. It is also the reason this page cannot give you a ranking: a claim without a number cannot be placed next to a number.

Services that make an accuracy claim without publishing a figure
ServiceWhat it saysWhere we read it
Fireflies the home page calls it the industry leader in transcription accuracy, with no percentage attached fireflies.ai, home page, checked
Otter.ai the source does not report an accuracy figure on the pages we opened otter.ai/pricing, checked
Maestra the source does not report an accuracy figure on the pages we opened maestra.ai/pricing, checked

How to read a percentage

  • The word up to is doing the work. Up to 99.9% is a ceiling, not an average. Two services can both write up to 99% and behave very differently on the same noisy street recording.
  • The test is usually not published. None of the 6 figures above comes with the recording that was used, the number of words counted or the rule for deciding that a word was wrong.
  • The conditions are the useful part. Sonix names audio quality, background noise, speaker clarity and vocabulary. VEED names audio quality and speech clarity. Those are the four or two things you can check before you believe a number on your own material.
  • Automatic and human are different products. One of the figures above, the one guaranteed in 12 hours or less, belongs to a service where people do the work. It cannot sit in the same column as an engine without a note, so it does not.
  • Some pages carry two numbers. HappyScribe writes 99% accurate in its navigation and shows 93% for English in the chart on its home page. We wrote down both, because choosing the flattering one would not be honest.

What we publish instead of a percentage

Our pages describe the shape of the result rather than a score: one line per segment, with the time it starts and ends and the speaker it belongs to, and the export files carrying the same structure. We have not run a benchmark against these services and we have not found one that publishes a common test, so we are not going to write a percentage of our own next to theirs.

If you need to know how a service handles your material, the honest method is a small one. Take a recording you know well, a minute of it is enough to start, send the same minute through the services you are considering, and read the transcript against the recording you remember. That gives you a number for your audio, which is worth more than any figure printed on a front page.

What this page does not tell you

It does not rank the services, because no common test has been published for them to be ranked on. It does not repeat a figure that we did not read on the page of the service that owns it. And it does not treat the word up to as if it were the typical result. Every number in the table above was copied from the address in the last column, on the date in the same cell, and nothing was carried over from a marketing summary or from memory.

Try it now

Send a minute you know and read it against the recording.

The widget below is the same one as on the home page: file → transcript with timecodes and speakers → export. The transcript is saved to your account.

Drop an audio or video file

or — mp3, wav, m4a, mp4, up to 64 MB and two hours

No account yet? The free plan needs no card. Create a free account or sign in.

FAQ

Questions this page answers.

Which service publishes the highest accuracy figure?

The highest figure on this page is the up to 99.9% that VEED writes about its automatic transcription, and the highest for an engine that names a language is the 99.8% of TurboScribe. Both numbers come from the pages in the table, and both are written as ceilings rather than as the usual result.

Does Transcrib Si publish an accuracy percentage?

No. Our pages describe the structure of a transcript, one line per segment with its time and its speaker, and they publish no percentage of correct words. We did not write one on this page either, because we have no measurement of our own that we could show you.

Are the percentages in the table comparable with each other?

No, and we will not pretend otherwise. None of the services publishes the recording, the number of words or the rule for a wrong word, so the figures were produced under conditions that differ from service to service. The only fair reading of each row is the sentence it came from, together with the conditions printed next to it.

Which of these is human transcription?

One of them. The figure of 99%+, guaranteed in 12 hours or less, belongs to the human service of Rev and stands in the link to it on the home page. The other 5 figures in the table describe automatic engines, and they were read from the product pages of those engines.

How do I find out which one is right for my audio?

Take a minute of a recording you know by heart, send the same minute to each service, and read the transcript against the recording. That gives you a figure for your own material, which is worth more than any percentage printed on a front page, and it takes one minute rather than a reading of six price lists.

Keep reading

Related pages.

Speaker labels compared

Who separates the voices, and in what words.

Transcrib Si vs Maestra

A suite selling minutes by the pack.

Transcrib Si vs Otter

A meeting notetaker with its own seats.

Transcrib Si vs Notta

Another assistant built around the call.

Transcrib Si vs TurboScribe

The service behind the 99.8% in the table.

Transcrib Si vs Fireflies

The service that claims the lead without a number.

Transcribe Russian audio to text

What our engine does with Russian.

Transcribe audio to text

The general guide to the whole service.

Plans and prices

What a month of our service costs.

Written by a person

Mark Burmistrov, Founder.

Every number on this page comes from a public page we opened ourselves, from the service contract or from our own measurements. The author of the page and the publisher of the site are the same person. Questions and corrections: support@transcrib.si.