Most Accurate Free Arabic Transcription? Honest Answer

We could not find an independent, published head-to-head test of free Arabic transcription tools, so nobody can honestly name the most accurate one for every recording, and we will not pretend to. What can be checked is narrower: how a model's size changes its Arabic errors, which Arabic varieties the big vendors say they support, and how to run a fair test on your own audio in ten minutes. We make Scriptorium, one of the tools in this space, so read what we say about it with that in mind.

What we measured ourselves

On 2026-10-24 we ran three Whisper model sizes on 3 clips of Arabic (about 246 seconds): sentences read by Egyptian speakers from the FLEURS dataset, scored after unifying alef, yeh and teh marbuta forms and removing diacritics.

The size of the model mattered a lot: on these clips the smallest model made roughly twice as many word errors as small. That is three clips of read speech from one dialect group. It says nothing about Gulf, Levantine or Maghrebi speech, about phone calls, or about overlapping speakers, and we have not measured the medium or large models. The method and per-clip data are on the measurements page.

What the free preview uses

The no-account preview always uses the small model and takes one file at a time, up to 15 minutes and 25 MB, with 3 tries a day. That is a limit, not a boast: it is the middle of the sizes we measured, not the largest available.

Which Arabic varieties do vendors say they support?

Support is not accuracy, but it tells you what a service is built to recognise. Each row is from the vendor's own documentation, read on 2026-10-26.

These are developer services and platforms, mostly paid beyond a free allowance, not all free consumer websites, and we did not read their prices. A service listing more locales is not shown to be more accurate; it only means it lets you choose one. Scriptorium has one Arabic setting among 33 languages and does not offer a dialect choice (more in how to transcribe Arabic dialects).

Run a fair test in ten minutes

1. Cut 60 to 90 seconds from your own recording. Choose the hardest part you can accept, not a clean demo.
2. Type the correct text yourself, or fix one draft slowly while listening. This is your reference.
3. Upload the same clip to every tool you are comparing, setting the language to Arabic by hand.
4. Count the words each tool got wrong and divide by the number of reference words, normalising the same way for every tool (how to calculate word error rate).
5. Look at what kind of errors there are: names and numbers cost more than a missing particle.

What about dialects and mixed speech?

OpenAI's own model cards say Whisper performs unevenly across languages and varies by accent and dialect, and that performance tracks the amount of training data per language. That is a general warning, and it applies to Arabic dialects as much as anything. Arabic speakers who switch to English mid-sentence are harder again; read the dialect guide and our Arabic page تفريغ اللهجات العربية.

When no free tool is accurate enough

For a legal record, a published quote or a medical or financial decision, an automatic draft needs a human reviewer who speaks the dialect, whichever tool produced it. See automatic or human transcription (in Arabic) and, for Arabic-language tool comparisons, أفضل مواقع تحويل الصوت إلى نص بالعربي.