How accurate is our transcription?
We publish no single accuracy percentage, and here is why: a score without the test set behind it means nothing. What this page gives you is the method we would use to measure it — the corpus, the sample, the normalisation, the metric, the engine build and the date — plus the numbers we do have: how long each step of our own pipeline takes.
The measurements we do publish
Four numbers, all from our own API on 21 October 2026. The conditions: the public demo sample (a 5.8-second Russian clip, 93 KB), one run, no load test behind the numbers. They are individual measurements, not averages over many runs, and we do not hide that: a re-run of the same file on the same day measured the recognition call itself at 0.08 s and the diarization round trip at 0.64 s. The spread between the two sets is the difference between the whole request and one call to the engine, and the load the engine was under at that moment. Read them as the order of magnitude, not as a promise.
| Step | Our measurement | What it covers | Date |
|---|---|---|---|
| Transcribe | 0.84 s | the recognition of a short file, from the request to the answer | 21 October 2026 |
| Detect the language | 0.39 s | the automatic language detection, when the language is not set by hand | 21 October 2026 |
| Speaker labels | 0.50 s | reading the speaker result back | 21 October 2026 |
| Diarization | 10.8 s | the whole round trip: asking for the speakers and waiting for the answer | 21 October 2026 |
The accuracy benchmark we will publish
The method is published now. Our own numbers come after the measurement, and they will come with the run: the split, the file list, the date and the engine build. We are not going to print a percentage before the run. What the market has already published, on the other hand, we collect and read: the accuracy benchmark lists the WER figures of AssemblyAI, Deepgram, Sonix, Sber's GigaAM and Whisper large-v3, each row with its corpus, its source link and the date we read it.
| Part of the method | What it will be |
|---|---|
| Corpus | a public Russian speech corpus (Golos) with a fixed test split, plus a set of our own real recordings of meetings, lectures and interviews |
| Sample size | stated with the result; the split and the list of files are published with it |
| Normalisation | lower case, punctuation removed, numbers and product names kept as they were spoken |
| Metric | WER (word error rate) and CER (character error rate), both standard for Russian |
| Engine build | the build of our own engine on the day of the run, dated |
| Conditions | the hardware, the language setting (fixed or automatic) and whether the speaker separation was on |
| Date | the day of the run, published with the number |
What happens to your file
The pipeline is the same for every file, and each step is measurable:
- The file is decoded. Whatever the container — mp3, m4a, wav, mp4 — the audio is unpacked into plain audio before the recognition pass.
- The language is decided. Either you set it (
ruoren) or it is detected automatically; the answer says which source was used: user, auto or fallback. The detection took 0.39 s in our measurement. - The recognition pass runs. It returns the text and the segments. The engine works in 22-second fragments with a 3-second overlap, so a long file is processed in pieces and the overlaps are stitched back together. A short file was transcribed in 0.84 s in our measurement.
- The speakers are separated, if you asked for it. That is a separate pass, and it is the slow one: 10.8 s for the whole round trip in our measurement. The words do not change — the labels are added to the segments.
- The result is written with the times. Every segment carries its start and its end, and that is what the exports carry as well.
What we will not do
- We will not publish a percentage without the test set behind it. Ask for the evaluation report and the sample set, and we will send both.
- We will not name the vendor of the engine in our copy, and we will not compare accuracy numbers we have not measured ourselves.
- We will not claim that we are the most accurate service on the market. The fastest way to judge us is to put your own file on the widget: three files a day, free, without a card.
Put a file on the page and read the first lines.
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
Transcription runs under your account: sign in or create a free account.
No account yet? The free plan needs no card. Create a free account or sign in.
Questions this page answers.
What is WER?
The word error rate: the number of wrong, missing and extra words divided by the number of words in the reference transcription. CER does the same with characters. Both are standard, and both need a reference text to compare against.
Do you publish a WER for Russian?
Not yet. The method is published above, and the numbers come after the run, with the date, the split and the list of files. We would rather publish a number late than publish a number we cannot defend.
Is your model the same one the big services use?
We do not publish the vendor of the engine, and we do not compare numbers we have not measured ourselves. What we publish is how our own engine behaves and how the pipeline is built.
How do I judge the quality for my own audio?
Upload a file whose content you know and compare the text with it. That is a harder test than any percentage, and it is the one that matters for your work.
Does the speaker separation change the words?
No. The labels are added to the segments; the text stays the same.
Which engine build are the numbers on this page from?
Our own engine, a model profile tuned for Russian and for the mixed Russian-English speech, as it was on 21 October 2026. We date every number we publish.
Related pages.
Transcribe Russian audio to text
What our own engine does with Russian and with mixed speech.
The general guide: formats, limits, exports and prices.
The same engine from your own code, with the endpoints and the error codes.
Every plan, every limit, in one table.
Upload with drag and drop, the queue and the exports.
The live answer base: 99 questions about the price, the limits, the languages, the exports, the privacy and the API, each with the page of this site it comes from.
Mark Vesyoly, Founder.
Every number on this page comes from the service contract, the pricing page or our own measurements. The author of the page and the publisher of the site are the same person. Questions and corrections: support@transcrib.si.