Word error rate (WER) is the standard measure of speech-to-text accuracy. It is the percentage of words a system gets wrong when its transcript is compared against a correct human reference. Lower is better: a WER of 10 percent means one word in ten is wrong.
How WER is calculated
WER counts three kinds of mistake against the reference transcript: substitutions (one word heard as another), insertions (a word added that was never said), and deletions (a word dropped). Sum the three and divide by the number of words in the reference, N, so WER equals (S + I + D) / N.
A 100-word reference with 3 substitutions, 1 insertion, and 1 deletion scores a WER of 5 percent. The math is Levenshtein distance applied at the word level rather than the character level.
Because insertions count, WER can exceed 100 percent. The mirror metric word accuracy is simply 100 percent minus WER.
WER scoring conventions
A raw word-for-word comparison is unfair unless both transcripts are normalized first, and this is where scoring conventions matter. The long-standing reference is NIST's Speech Recognition Scoring Toolkit and its sclite tool, which fixed how alignment and error counting are done for benchmark evaluations.
Normalization decides whether "twenty five" and "25" count as a match, how punctuation and casing are treated, and whether contractions get expanded. Two providers can report different WER on the same audio purely because they normalized differently, so a stated number means little without the scoring recipe behind it.
What counts as a good WER
A good WER depends heavily on the audio. On clean, well-recorded speech, leading models land in the 5 to 10 percent range, and the strongest sit below 5 percent on easy benchmarks.
Real conditions push it higher: background noise, crosstalk, strong accents, phone-quality audio, and specialized vocabulary all degrade it. For finer comparison, some teams also track character error rate, which is more forgiving of small spelling slips.
A hallucination, where the model invents fluent text with no audio behind it, is especially costly. It reads as confident and correct, so it inflates real-world error in ways a quick skim of the transcript misses.
How to use WER when comparing providers
Treat published WER numbers with suspicion. Vendors quote results on clean benchmark datasets that rarely resemble your audio, and each may use its own test set and normalization, so the figures are not comparable across providers.
Independent efforts like the Open ASR Leaderboard help by fixing the data and scoring. Still, the only number that matters is WER on a sample of your own recordings, transcribed by each candidate and scored against a reference you trust.
Run that test before committing, because a two-point WER gap on a benchmark can flip once real audio is involved.
Frequently asked questions
How is word error rate calculated? Add the substitutions, insertions, and deletions between the system transcript and a human reference, then divide by the number of words in the reference. The result is usually shown as a percentage, and lower is better.
What is a good word error rate? On clean audio, 5 to 10 percent is typical for strong models, and under 5 percent is excellent. Noisy calls, heavy accents, and jargon push it higher, so a good WER depends on the audio and the domain.
Can word error rate be over 100 percent? Yes. Because inserted words count as errors, a system that adds many spurious words can score above 100 percent, which also means word accuracy (100 percent minus WER) can go negative.