How to evaluate transcription accuracy claims
·
Verbafy Team · 2 min read
Many transcription products advertise 99% or higher accuracy. That number is only useful when you can check the model, dataset, normalization, and who measured it.
What Benchmark Word Accuracy means
Benchmark Word Accuracy is 100% − WER for an exact speech-to-text model on a named benchmark. It is evidence about that model under those conditions. It is not a guarantee for every Transcript.
How Verbafy reports accuracy
Verbafy currently uses AssemblyAI. On the independent Coval Speech-to-Text Benchmarks (30-day window), the AssemblyAI result Verbafy reports is 1.8% WER. Benchmark Word Accuracy is 98.2% (100% − WER). AssemblyAI is the best-in-class model family on that snapshot. The Coval leaderboard changes as new results arrive. Last checked on 4 October 2026. Coval is a third-party benchmark, so Verbafy reports a number someone else measured, not a self-reported claim. Real results vary with audio quality, accents, terminology, crosstalk, and speaker similarity.
Check any accuracy claim
| Check | What to look for | Verbafy's figure |
|---|---|---|
| Processor | Who transcribes the audio | AssemblyAI |
| Benchmark | A named, public test set or leaderboard | Coval Speech-to-Text Benchmarks, 30-day window |
| Metric | How the number is calculated | 100% − WER |
| Measurer | Who ran the test, and whether they sell the product | Coval, a third party |
| Date | When the number was last measured | Last checked on 4 October 2026 |
This evaluation method applies to YouTube to text, Spotify to text, and meeting to text.
Why a lower checkable number can be more useful
A lower, independently measured number is more useful than a 99%+ marketing claim that does not name a dataset, model, or measurer. The confidence is in evidence you can inspect. It is not a ranking that says Verbafy transcribes more words correctly than a 99% claim.
If a product names an independent benchmark, model, dataset, and date, treat that evidence with the same care.
What to check in a Transcript
Accuracy copy is not a substitute for playback. Open the timestamp and hear the original audio before you quote a name, number, or decision.