Accuracy
Instead of asking you to trust a marketing number, we measure it. We run our engine over public speech datasets with human reference transcripts, score the word error rate, and publish the results by recording condition. We re-run it monthly so it never goes stale.
Last run August 2026. Engine: ElevenLabs Scribe v2.
Run on the benchmark harness we share with FileToText: the same engine, the same frozen clips, the same scoring. The multi-speaker meeting set of that harness is not included, because meeting recordings are not YouTube videos.
Accuracy is highest on clean speech and drops on noise and crosstalk. We publish the hard rows too, because that is what makes the clean-audio number believable.
| Condition | Dataset | Accuracy | WER |
|---|---|---|---|
| clean | LibriSpeech test-clean | 98.5% | 1.5% |
| harder audio | LibriSpeech test-other | 97.1% | 2.9% |
| accents | Common Voice | 94.3% | 5.7% |
| talks | VoxPopuli | 92.7% | 7.3% |
Overall accuracy across the set: 95.7% (WER 4.3%).
These numbers apply to the API and the MCP server as well. A job queued from your own code, from an assistant, or from the paste box on this site runs through the same pipeline on the same audio, so there is no faster-but-rougher tier and no separate quality to pick. The developer docs cover how to queue one.
Word Error Rate (WER) is the standard way to score a transcript: the share of words that are substituted, inserted, or deleted versus a correct reference transcript. Accuracy is simply 100% minus the WER. Lower WER is better.
We use recognized public speech datasets, each with a human-verified reference transcript, so the test is reproducible and covers a range of real conditions:
Before scoring, both the reference and our transcript are put through the standard Whisper EnglishTextNormalizer, audio event tags removed (lowercasing, punctuation and number normalization) so we score real recognition errors, not formatting differences. The edit distance is computed with jiwer, the standard open-source WER tool. The engine under test is ElevenLabs Scribe v2. Everything here is reproducible: same datasets, same normalizer, same tool.
Clean, single-speaker audio transcribes at very high accuracy. Noise, heavy accents, and several people talking at once are harder, for every engine. That is why the editor shows a per-word confidence view that flags the words worth a second look, so you know exactly where to spend a quick review instead of re-reading the whole thing. Try it on a video of your own: the first 10 minutes are free.