Skip to content
Zahvox
GlossaryContent draft

Word error rate

Word error rate — WER — is the standard way to measure how accurate a speech-to-text system is. Lower is better.

Try the free voice toolFree · No signup · Runs in your browser

Definition

Word error rate is calculated by aligning the recognizer's output against a reference transcript and counting substitutions, insertions, and deletions, divided by the number of words in the reference. A WER of 10% means roughly one in ten words is wrong.

In plain language

Take what the recognizer wrote. Compare it to what was actually said. Count the mistakes. Divide by total words. That's WER.

Why it matters

WER lets you compare recognizers on the same footing. But it hides important detail: a 5% WER on clean American English studio audio is very different from a 5% WER on noisy conference audio with mixed accents. And WER doesn't measure how usable the output is — punctuation, casing, and formatting all matter for real writing workflows.

How this relates to Zahvox

Zahvox doesn't publish a WER for the free tool because its recognizer is your browser's Web Speech API, which varies by browser and version. What matters more for a voice-to-writing product is how easy it is to fix the last-mile errors in the editable transcript.

Status: General industry concept, not tied to a specific Zahvox release.

Example

If you say "schedule a call with Alex at three" (7 words) and the recognizer writes "schedule call at three", that's 3 errors out of 7 words — a WER of ~43% for that utterance.

Common misconception

A tool with lower advertised WER is not automatically better for your writing. Real-world accuracy depends on your voice, environment, and the vocabulary you use.

Frequently asked questions

Keep exploring

Improve voice-to-text accuracy
Voice-to-text not working
Glossary

Stop typing. Start thinking.

Open the free voice-to-text tool and turn your next thought into clean text in seconds.

Open the free tool