← AI Terminology
WER - Word Error Rate
Word error rate (WER) is the edit-distance error rate between recognised speech transcripts and reference text (insertions, deletions, substitutions).
It is the primary metric for automatic speech recognition.
It is the primary metric for automatic speech recognition.
Why It Matters in AI
ASR systems live or die by WER on target domains (accents, noise, medical terms). Despite quirks (all errors weighted equally), it remains the industry standard for speech-to-text quality.
Key Points
| Aspect | Description |
|---|---|
| Use | Whisper and ASR model cards |
| Limits | Ignores semantic equivalence; punctuation variants |
| Formula | (S+D+I)/N reference words |
| Related | CER (character error rate), BLEU for MT |
| Practice | Domain WER matters more than clean LibriSpeech alone |
| Lower better | 0 = perfect match |
Simple Analogy
Grading a dictation by counting how many word fixes a teacher needs with a red pen to match the script.
Common Usage Examples
- Compute WER with
jiwer - Report per-domain WER
- Compare Whisper model sizes
- Augment audio to stress-test WER
Summary
In short: WER is the edit-distance error rate for speech transcripts — the standard yardstick for speech recognition accuracy.