Choosing transcription software comes down to four questions: whether the recording may leave your computer, whether you need live or file-based transcription, what language the speakers use, and whether the result needs legal standing. Browser tools like skryba.ai cover the most common case — an existing file, remote work, text for further editing.
Upload a recording. Create an account only after you see the result.
Drop a file here or choose one from your device. We will show you the beginning of the transcript first.
Which transcription model gets the fewest words wrong?
skryba.ai transcribes on Scribe v2, which gets 2.2% of words wrong in Artificial Analysis' independent index — the lowest of every model measured. Gemini 3.5 Transcribe gets 2.6% wrong, Whisper large-v3 4.1%. It is the same model that writes the transcript preview you see before paying.
Word Error Rate — words transcribed incorrectly — lower is better
Model
WER
skryba.ai (Scribe v2)the model we transcribe on
2.2%
MAI-Transcribe-1.5Microsoft Azure
2.4%
Gemini 3.5 TranscribeGoogle
2.6%
Universal-3 ProAssemblyAI
3.1%
GPT TranscribeOpenAI
3.3%
Whisper large-v3OpenAI
4.1%
Nova-3Deepgram
5.2%
Source: Artificial Analysis, AA-WER v2 index, non-streaming mode — read 27 August 2026. Seven of the 40-plus models measured: the top of the table and the names you already know. The index is measured mostly on English audio; for other languages, Polish included, no comparable public ranking exists. That is why we show you the start of every transcript before we ask for anything.
Five situations, five different answers
Situation
The right category of tool
I have a recording and want text to edit
A browser tool — including this one
The recording may not leave my computer
Desktop software running locally
I want meeting notes, during the meeting
A bot that joins the call
I want to dictate and see text immediately
Dictation built into your operating system
The transcript needs legal standing
A provider the receiving institution accepts
Scroll the table sideways to see the remaining columns.
Only the first row leads to skryba.ai. The other four lead elsewhere, and we say so plainly, because the expensive mistake when choosing a tool is buying the right product for the wrong job.
The last row needs care: the requirements depend on the institution and the matter. A court record is produced by a court reporter, a sworn translator certifies translation between languages, and some authorities accept a certified-accuracy transcript from a specialist provider. Before ordering anything, ask the body that has to accept it what form it expects.
Browser or installed software
This is the most common decision, and it almost always reduces to one thing: whether the recording may be sent to somebody else's server. The rest is secondary.
Criterion
Browser
Desktop
Recording leaves your computer
Yes
No
Needs powerful hardware
No
Yes, especially for longer recordings
Time to first result
Minutes
Install, model, configuration
Working across devices
Yes
No
Cost
Per hour used
One-off or free, but your own hardware
Scroll the table sideways to see the remaining columns.
If confidentiality is the deciding factor, do not look for a cloud compromise — take a tool that runs locally. If confidentiality is not a blocker, a browser tool wins on time to first result and on not tying you to one machine.
What to compare
Accuracy in your language, tested on your recording — not on the vendor's demo sample. A tool without a free preview makes that test hard on purpose.
Speaker recognition, if you record conversations rather than monologues.
Timestamps and SRT export, if you produce subtitles.
The billing model: a subscription with expiring hours or a pack with none — depending on whether you record regularly.
What happens to the recording afterwards, and whether you can delete it yourself.
Where the data physically sits, if GDPR applies to you.
A claimed “99% accuracy” means nothing without stated conditions. Every model struggles with overlapping voices, a distant microphone and reverb — which is why the only honest test is your own recording, ideally the hardest one you have.
Yes — open-source models run locally can be very good and send the recording nowhere. The cost is setup, processing time on your own hardware and no editor. It is a sensible choice when confidentiality matters more than time.
My recordings are confidential. What should I use?
If the material may not leave your computer, choose a tool that runs locally — no cloud service satisfies that condition. If cloud is acceptable, check where the data sits, who can access it and whether you can delete it yourself.
Does skryba.ai join meetings and take notes?
No. We have no bot that joins calls. Record the meeting with your platform's own tool, then upload the file — you get a transcript with timestamps and speaker labels.
How do I compare accuracy across tools?
Take one recording — the hardest you have, ideally with several speakers and background noise — and run it through each tool. Compare the same material rather than vendor samples, and count errors on names and terminology, because those cost the most time to fix.
