How accurate Echo is in Spanish
We measured it with a public set of recordings so anyone can repeat it. Here are the result, the method and what the figure doesn’t tell you.
97.4%
of words right in Spanish
A word error rate (WER) of 2.63% with the model Echo uses. Measured on 8 November 2026.
Models we tested
| Model | Word error rate (WER) | Speed | Use |
|---|---|---|---|
| Whisper large-v3 turbo (Echo’s) | 2.63% | 1.4 times faster than the audio | The one we use: just as good in practice, and faster |
| Whisper large-v3, simple search | 2.53% | 1.0 times | Slightly more accurate and slower |
| Whisper large-v3, search of 5 | 2.33% | 0.9 times | The most accurate we tested, and the slowest |
How we measured it
- The public FLEURS dataset, in its Spanish version (es_419, Latin American Spanish: the one FLEURS publishes): 80 sentences read by different people, with their correct text.
- Before comparing, both versions are lower-cased, stripped of punctuation and numbers are written the same way, to measure words rather than formatting.
- WER = words changed, removed or added, divided by the words in the correct text. Accuracy = 100% − WER.
- The test lives in Echo’s code (bench/wer_bench.py) and is repeated whenever the model or settings change.
What the figure doesn’t tell you
- These are read sentences with good sound. In a real meeting, with noise, people talking over each other or a poor connection, accuracy is lower.
- 80 sentences is a small sample: it’s good for comparing models, not a guarantee.
- The remaining errors are mostly unusual proper names and how times and figures are written. For names there’s the glossary: add your industry’s words and Echo spells them right.
- We haven’t measured other languages yet, so we don’t give figures for them.
Who said what
To separate speakers, we measured conversations built from known voices (LibriSpeech): merging the same person split into several raised accuracy from 80.7% to 85.0% without mixing anyone up. In the app you can fix who said each paragraph in a moment.
Try it with your next meeting.
Create your account in a minute or ask us for a demo for your team.
