Can word error rate be greater than 100%?

WER has no upper limit, and insertions are what push it past 100 percent

Word error rate is (S + D + I) divided by N, the number of words in the reference. Substitutions and deletions cannot total more than N, so on their own they keep the score at or below 100 percent. Insertions have no such limit. Each extra word the recognizer writes adds one to the numerator and nothing to the denominator, so a system that inserts a lot of text can score far above 100 percent.

Morris and colleagues give a concrete case. A system that outputs two wrong words for each input word has no correct words, and its WER is 200 percent ¹. Their point is that both systems in the example communicate exactly zero information, yet the score for the second can be far above 100 percent. For how the score is calculated in the first place, see the question on calculating WER.

A worked example scores 200 percent with four insertions

Take a two-word reference, "one two", and an output of six words, "one two three four five six". The two correct words are hits. The four extra words are insertions. Nothing is substituted or deleted.

In our test, the jiwer package returned 2.0 for this pair, and it counted four insertions. The arithmetic is the same whatever the wording of the extra words.

Insertions dominate some real errors, so the score can climb quickly

Insertions can dominate for some speakers. In the second phase of an Apple study of speech from people who stutter, 80.9 percent of the recognizer's errors were extra inserted words, and only 1.6 percent were missed words ². Insertions of that kind are where a score can climb fastest.

Hallucination is the most extreme source. In a 2024 study of Whisper, roughly 1 percent of audio transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ³. A single invented sentence can add more insertions than a short clip has reference words. A 2025 study found that when non-speech sounds were played to Whisper, a set of hallucinated phrases appeared frequently ⁴. Those extra words can push a short clip's WER far past 100 percent.

Empty references are a separate case with their own rule

The formula has no value when the reference is empty, because N is zero. The jiwer README calls this undefined behaviour and notes that it divides by zero ⁵. Since version 4.0, jiwer defines the case: when the hypothesis is non-empty, every word counts as an insertion ⁶. In our test, jiwer.wer("", "peaceful silence") returned 2, which is 200 percent for a clip that should have been silent.

This is useful for testing silence. A recognizer that writes text for a silent clip shows up as a non-zero score, and an empty output for an empty reference scores 0. Check that your scorer handles empty references on purpose, rather than letting it crash or quietly skip those clips.

Accuracy goes negative when WER is above 100 percent

Accuracy is one minus WER, so a WER of 200 percent gives an accuracy of -1.0. That is a sign of trouble in the output, not a meaningful score. Clipping accuracy at zero would hide how much extra text the system wrote. If you report accuracy, report the insertion count next to it, so a reader can see whether the problem is substitutions or extra text. The question on WER and accuracy explains the relationship.

What to do when a score is above 100 percent

1. Read the three counts. Use the per-type counts, not only the total. A high insertion count points to repeats, extra sentences or hallucination. A high substitution count points to wrong words.
2. Look at the output, not the number. Find the clips with the most insertions and read them. Check for repeated phrases, sentences that the audio does not contain, and text written during silence.
3. Test on silence and on non-speech. Run clips with no speech, such as room noise, through the recognizer. Any words it writes are insertions.
4. Do not cap the score to hide the problem. Report the raw value and explain it. A capped number makes two different systems look the same.
5. Compare like with like. Scores above 100 percent from different tools can differ because of how they handle empty references and inserted text, so check the tool and version.

The guide to word error rate explains the other error types, the overview of speech recognition places the score in context, and the question on what a good score is covers how to set a target. For a Python version of the calculation, see how to calculate word error rate in Python.