What is Word error rate (WER)?

Why is WER important for evaluating speech recognition?

WER provides a standard way to compare a transcript with a reference, helping teams quantify substitutions, deletions, and insertions.

• Creates a repeatable baseline for ASR testing.
• Shows whether model changes improve transcription accuracy.
• Supports comparison across languages, accents, and audio conditions.
• Helps locate recognition weaknesses before deployment.

How Word error rate (WER) works

• Align the reference transcript and recognition hypothesis using the minimum sequence of substitutions, deletions, and insertions.
• Add those word errors and divide by the number of words in the reference.
• Aggregate counts across an evaluation set rather than averaging short utterance percentages blindly.
• Segment the result by language, audio condition, and business importance.

Word error rate (WER) in action

A transcript changes “do not renew” to “renew.” It may contribute only one substitution to WER, but it is a critical semantic error, so the team tracks negation and action entities separately.

How Word error rate (WER) is measured or evaluated

Use representative production conditions and explicit definitions. Review these measures together:

• WER: substitutions plus deletions plus insertions divided by reference words
• substitution, deletion, and insertion rates
• critical-entity error rate for names, amounts, dates, and negation
• WER by language, codec, noise, speaker, and intent

Limitations, risks, and common failure modes

• Reference transcripts can contain their own mistakes.
• Tokenization and normalization rules change the score.
• Low WER can coexist with harmful entity errors.
• Short utterances produce volatile percentages.

Word error rate (WER) vs. character error rate

WER scores edits at the word level. Character error rate scores character edits and can be more informative for languages or tasks where word segmentation differs.

What teams should consider when using or implementing Word error rate (WER)

• Publish normalization and scoring rules.
• Use reviewed reference transcripts.
• Report error counts and confidence intervals.
• Pair WER with entity, semantic, latency, and task metrics.