Skip to main content
OT

Word error rate (WER)

Also known as：WER · character error rate · CER

Word error rate (WER) is the standard measure of speech recognition accuracy, equal to the number of insertions, deletions and substitutions divided by the total words in a human-labelled transcript. Chinese normally uses character error rate (CER) instead.

The formula

WER = (I + D + S) / N × 100
  • Insertion (I) — words in the output that were never spoken;
  • Deletion (D) — words in the reference that the output dropped;
  • Substitution (S) — words replaced by different ones;
  • N — total words in the human-labelled reference transcript.

This is the definition Microsoft uses in its custom speech evaluation documentation, which also points at sclite from the NIST Scoring Toolkit (SCTK) for reproducing the calculation locally.

Chinese uses character error rate

Chinese has no natural word boundaries, so different tokenisers produce different word error rates from the same output. Chinese work therefore uses character error rate (CER), which swaps the unit from words to characters and leaves the formula and the three error categories unchanged.

The error mix points at different problems

Two tests with the same score can have entirely different causes. Microsoft documents the mapping:

  • Deletion-heavy results usually mean weak audio signal strength, so capture needs to move closer;
  • Insertion-heavy results mean a noisy recording environment, possibly with crosstalk;
  • Substitution-heavy results mean insufficient coverage of domain-specific terms.

Read the split before deciding whether the next fix is hardware or glossary.

What it does not cover

Word error rate compares words only. Punctuation, capitalisation and number formatting are outside it. Microsoft defines an extended metric, token error rate (TER), that folds in punctuation, capitalisation and inverse text normalisation. Caption readability tracks TER far more closely than it tracks WER.

Numbers only compare under identical conditions

Published vendor accuracy figures are almost always measured on clean audio with standard pronunciation. The only figure worth acting on is the one you measure on your own broadcast audio under fixed conditions — change the background music, the microphone or the speaking rate and the numbers stop being comparable.

Related terms