Fermion ResearchFermion Research

Glossary/facts checked 13 November 2026

Word error rate (WER)

The standard accuracy number for speech recognition, how to compute it by hand, and the details that make two WER figures comparable or not.

Phonon-2.5 benchmarks and downloadsOpen weights on Hugging Face.

The short answer

Word error rate (WER) is the share of words a speech recognition system gets wrong: the number of substituted, deleted and inserted words, divided by the number of words in the reference transcript. Lower is better. A WER of 5 % means about one error for every 20 reference words.

The formula is WER = (S + D + I) / N, where S, D and I are counted from the minimum edit alignment between the reference and the model's output, and N is the number of reference words. Because insertions are counted, WER can go above 100 %.

WER, CER and sentence error rate

Three error rates appear in speech recognition papers and leaderboards. They answer different questions, so a number is only meaningful next to its name.

MetricWhat it countsUnit of the editTypical useWatch out for
Word error rate (WER)Substituted, deleted and inserted words over reference wordsWordMost languages written with spacesDepends on text normalisation; one wrong suffix costs a whole word
Character error rate (CER)Substituted, deleted and inserted characters over reference charactersCharacterLanguages written without spaces between words, such as Chinese, Japanese and ThaiUsually much lower than WER on the same audio; never compare a CER with a WER
Sentence error rate (SER)Utterances with at least one error over all utterancesWhole utteranceCommands and short prompts, where any error fails the requestSays nothing about how many words were wrong

A worked example

Reference (8 words): please send the report to Ada by Friday. Model output: please send a report to Ada Friday today.

  • Substitution: "the" became "a" (S = 1).
  • Deletion: "by" is missing (D = 1).
  • Insertion: "today" was added (I = 1).

WER = (1 + 1 + 1) / 8 = 37.5 %. Six of the eight reference words came through correctly, so only two reference words were lost, but the inserted word also counts: three errors, not two.

Across a test set, WER is computed on the totals: all errors over all reference words. That is not the same as averaging the WER of each utterance, which gives short utterances far more weight. Libraries such as jiwer compute the corpus form.

Why two WER numbers may not be comparable

  • Normalisation: before counting, both texts are usually lowercased and stripped of punctuation, and numbers and spellings are folded to one form. "9:30" against "nine thirty" is two errors without a normaliser that maps both to the same text. See text normalisation for ASR.
  • The test set: the same model scores very differently on different audio. Phonon-2.5 scores 1.60 % on LibriSpeech test-clean (read audiobooks) and 9.12 % on AMI (meeting recordings), scored with the same code.
  • The average: an average over seven sets is not comparable with an average over nine, or with a mean over languages. A good results table says which sets an average covers.
  • The subset: a sample of a test set gives a noisier number than the full set. Results on full test sets, scored with one code for every model in the table, are the ones to compare.

Questions

What is the formula for word error rate?
WER = (S + D + I) / N: substitutions plus deletions plus insertions, divided by the number of words in the reference transcript. The counts come from the minimum edit alignment between reference and output.
Can word error rate be more than 100 %?
Yes. Insertions are counted but the denominator is only the reference words, so a model that adds many words that were never said can exceed 100 %.
What is a good word error rate?
It depends on the audio. The same model can score under 2 % on clean read speech and around 9 % on meeting recordings, so compare models on the same test sets rather than against a fixed threshold.
What is the difference between WER and CER?
WER counts errors per word; CER counts errors per character. CER is used for languages written without spaces between words, such as Chinese, and the two numbers should never be compared with each other.
Is word accuracy just 100 % minus WER?
Loosely, yes, and some reports quote it that way. Because WER can exceed 100 %, 100 minus WER can go negative, which is why most benchmarks report WER directly.
Should WER be averaged per utterance or over the whole test set?
Standard practice is over the whole set: total errors divided by total reference words. Averaging per-utterance WER gives short utterances too much weight.

Related

Sources and method

  • Phonon numbers: Phonon-2.5 release post (/research/phonon-2-5/), 2026-10-22; Fermion Research measurement, full test sets, one scoring code for every model in a table.
  • Open ASR Leaderboard: github.com/huggingface/open_asr_leaderboard README, read 2026-11-13.

Canonical: https://www.fermionresearch.com/glossary/word-error-rate/