Word Error Rate Calculator
The word error rate of a speech recogniser or transcript from its substitutions, deletions and insertions against the reference — with word accuracy, the error mix, and what the rate means per hour of speech.
Align the transcript with the reference and count three kinds of error: words swapped, words missed, words added.
How the word error rate calculator works
Align the transcript with the reference and count three kinds of error: words swapped, words missed, words added. WER is their sum over the words in the reference. It can exceed 100%, because insertions are counted too.
A good broadcast-speech system is under 5%; conversational, accented or noisy audio runs 10 to 20; a rate above 30 is hard to read. The error mix tells you what to fix: deletions point at dropped audio, insertions at hallucination.
Formula: WER = (S + D + I) / N
Worked examples
| Inputs | Word error rate | Note |
|---|---|---|
| A 7% transcript | 7% | good |
| Noisy audio | 23.333% | 23% |
| Hallucinating | 13.75% | insertions dominate |
FAQFrequently asked questions
What is word error rate?
The number of word-level substitutions, deletions and insertions needed to turn the transcript into the reference, divided by the reference length. It is the standard speech-recognition metric.
Can WER be over 100%?
Yes — if the system inserts many words that were never said, errors can exceed the reference length. It is a rate of edits, not a share of words wrong.
What is a good WER?
Under 5% on clean read speech, 10 to 15% on conversational audio, and anything under 20% is usable with light editing. Human transcribers score around 4 to 5%.
How are the three error types found?
By aligning the two word sequences with minimum edit distance — the same dynamic programme as Levenshtein distance, on words instead of characters.
Does normalisation matter?
Enormously. Casing, punctuation, numerals versus words and contractions can move WER by several points. Normalise both sides the same way before counting.
Where these figures come from
- Vaswani et al. (2017) — Attention Is All You Need — the transformer architecture the memory arithmetic follows
- Kaplan et al. (2020) — Scaling Laws for Neural Language Models — the compute relationship used for training estimates
- Hoffmann et al. (2022) — Training Compute-Optimal Large Language Models — the tokens-per-parameter guidance ("Chinchilla")
- IEEE 754 — Standard for Floating-Point Arithmetic — the numeric formats behind bytes per parameter
- National AI Centre — Australia's national AI body
Last checked: September 2026. The relationships here are architectural, not vendor-specific: bytes per parameter follow the numeric format, KV-cache size follows the transformer definition, and token-per-word ratios come from published tokeniser behaviour.