Work out word error rate instantly with clear inputs, formula shown and shareable results.
Word error rate is the Levenshtein distance between the recognised and reference word sequences, normalised by the reference length. Because insertions are counted but do not lengthen the denominator, WER can exceed 100 percent when a system hallucinates extra words. Aligning the hypothesis to the reference first is what produces the substitution, deletion and insertion counts.
Word error rate
WER = (S + D + I) / N
Yes. A system that inserts many spurious words can accumulate more errors than there are reference words, which is why accuracy is reported as 1 - WER and may go negative.
Apply consistent casing, expand or remove punctuation, and standardise numbers and contractions. Comparing WER figures produced under different normalisation rules is meaningless.