Word Error Rate (WER): How to Calculate and Read It

Word error rate is the share of spoken words a recognizer got wrong, measured against a human reference transcript

Word error rate (WER) is the standard score for automatic speech recognition (ASR), the software that turns speech into text. It compares the machine's output with a reference transcript, a text that a careful person wrote from the same audio, and counts the words that differ. Lower is better, and a score of 0 means the two texts match word for word. Morris and colleagues describe WER as the de facto standard index for ASR assessment, defined as the proportion of word errors to words processed ¹.

A WER of 5 percent means about one word in twenty is wrong. That is a useful rule of thumb, but it hides what kind of errors they are and where they fall. A 2026 study notes that ASR reaches word error rates below 5 percent on typical speech ². For dysarthric speakers, people whose speech is affected by a motor disorder, the error rates are much higher, as the figures below show. The same number can describe a clean read passage and a noisy phone call, so the rest of this guide shows how to calculate WER, how to read it and when it misleads. For the broader picture, see our overview of speech recognition and the guide on how speech recognition works.

WER counts three kinds of mistakes: substitutions, deletions and insertions

Every difference between the reference and the output falls into one of three types. Correct words are called hits, and they are not counted as errors.

People tend to think of recognition errors as wrong words, which means substitutions. Insertions are the type most often overlooked, and they can dominate for some speakers. In 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 ³. A word that was never said can be as damaging to a reader as a wrong word, and WER counts both the same way. The difference between the three types matters when you decide what to fix. For everyday fixes, see dictation mistakes.

The formula is (S + D + I) divided by N, the number of words in the reference

The formula is short:

WER = (S + D + I) / N

Here S, D and I are the counts of substitutions, deletions and insertions, and N is the number of words in the reference transcript. Koenecke and colleagues define the denominator the same way, as the total number of words in the correct transcript ⁴. The denominator is the reference, not the output, which is why the score is fixed by the human transcript.

Here is a worked example with a made-up sentence. The reference has 10 words, and the recognizer made one substitution, one deletion and one insertion.

The arithmetic is the whole method. We checked this pair with jiwer, an open-source Python package, and it returns the same 0.30. The page on how to calculate word error rate in Python shows how to run the same check on your own transcripts, and how to calculate word error rate walks through the steps by hand.

WER can also go above 100 percent, and that surprises many readers. Because insertions go in the numerator while the denominator is the reference length, a system that writes a lot of extra text can score more than 100 percent. Morris and colleagues give an example of a system that outputs two wrong words for each input word, which they score as a WER of 200 percent ⁵. In that example, both systems, the 200 percent one and a system that gets every word wrong, communicate exactly zero information ⁵. The question on whether WER can exceed 100 percent covers the edge cases.

Accuracy is the same arithmetic in reverse. Word accuracy is (N - S - D - I) / N, which equals 1 minus WER. It therefore falls below zero when insertions are numerous. Is word error rate the same as accuracy? explains the difference in more detail.

The alignment behind WER is an edit distance computed on words, not letters

The scorer cannot count errors until it decides which output word corresponds to which reference word. That is a sequence-alignment problem, solved with dynamic programming: the scorer finds the set of edits that turns the output into the reference with the fewest changes. This is the Levenshtein edit distance, applied to words instead of letters. The jiwer package describes its measures the same way, as computed "with the use of the minimum-edit distance between one or more reference and hypothesis sentences" ⁶.

The weights used in that search are where tools differ. Morris and colleagues show that the usual procedure, with zero weight for hits and equal weight for substitutions, deletions and insertions, is optimal under their assumptions ⁷. The NIST sclite scorer, which many research papers use, weights the alignment differently. Its documentation (version 1.2) says it weights correct words, insertions, deletions and substitutions as 0, 3, 3 and 4 respectively ⁸. Weights decide which alignment wins when several have similar edit counts, so they can change how errors are split between the three types, even when the totals are close. Write down the tool and version with every score you report.

The same arithmetic works on characters. In the 2024 Mandarin stuttering challenge, character error rate (CER) was defined as the minimum number of insertions, substitutions and deletions needed to turn the recognizer output into the reference ⁹. Mandarin has no spaces between words, so counting words first needs a segmentation step, which is one reason CER is common for Chinese. The comparison of WER and character error rate covers when to use each.

Text normalization (case, punctuation, numbers, spelling) can change a score, and numbers are one of the traps

Before scoring, most leaderboards and papers normalize both texts. They lower the case, strip punctuation and often remove filler words. The Open ASR Leaderboard does this for its normalized transcripts, removing casing and punctuation, along with disfluencies such as fillers, before computing WER ¹⁰. That is a fair way to rank systems on one rule set, but a normalized score says nothing about whether your output has the right punctuation or keeps the "um" and "uh" you may want to remove.

Numbers are one of the traps. A recognizer may write "one hundred" as "100", while the reference uses words. Kacprzak and Fraś measured how much WER changes when numbers are left unnormalized in Polish and VoxPopuli speech, and found that the gap may be substantial, more than 2 percentage points on some datasets ¹¹. Their conclusion is that the lack of number normalization can produce a misleading system comparison ¹². Their study covers Polish and multilingual data, so the size of the effect in English may differ.

Spelling variants (British and American spellings, contractions, hyphenated compounds) work the same way: a correct word written differently is scored as a substitution unless the normalizer maps it. The practical rule is to write down the normalizer, apply the same one to the reference and the output, and never score a raw output against a normalized reference. If you need punctuation or verbatim output, the guide to verbatim and clean verbatim transcription explains the choices a transcript can make, and inverse text normalization is the step that turns spoken numbers into written ones.

A WER score describes one test set, not your voice or your microphone

A WER number belongs to a specific audio set, a specific speaker group and a specific recording setup. Change any of them and the score can move a lot. A Conformer model reached 2.1 percent on the clean LibriSpeech test set and 4.3 percent on the harder "other" set, without a language model ¹³. The same model, the same metric, two different scores. Both sets come from LibriSpeech.

Microphones matter as well. In a 2017 study, a recognizer scored a WER of 3.7 percent on close-talking sentences read in a studio ¹⁴. On real recordings from a single distant ceiling microphone in an apartment, the same kind of recognizer scored 12.0 percent ¹⁵. The two figures come from different test sets, so they show how much conditions can change a score, not the effect of the microphone alone. Your desk, room and headset belong in the test, not only the model.

Speakers matter too. A baseline recognizer had a WER of 3.4 percent on typical speakers and 36.3 percent on speech from people with Parkinson's disease ¹⁶. Fine-tuning on Speech Accessibility Project data brought the second figure down to 23.7 percent, which still leaves roughly one word in four wrong ¹⁶. For severe dysarthria, a 2026 benchmark found that every commercial system it tested exceeded 51 percent WER at its default settings ¹⁷. An accent can change the score in the same way, which is why our guide to speech recognition and accents recommends checking with your own voice.

So a published WER is a claim about one test. The useful question is whether the test resembles your audio. If it does not, record a short sample of your own, transcribe it carefully, and score that.

Human transcribers make errors too, so the human score sets a floor on hard audio

A reference transcript is itself made by people, and people make mistakes. That is why a very low WER on hard audio is hard to reach even for a human. In a 2017 IBM study, the best of three human transcribers reached a WER of 5.1 percent on the Switchboard telephone test after quality checks ¹⁸. On the harder CallHome telephone test, the best human transcriber reached 6.8 percent ¹⁹. The authors' recognizer reached 5.5 percent and 10.3 percent on the two subsets of the Hub5 2000 evaluation ^{20 21}.

On Switchboard the recognizer was close to the human transcriber. On CallHome the gap widened, from the best human's 6.8 percent ¹⁹ to the recognizer's 10.3 percent ²¹. Both sets are old telephone conversations, not dictation into a microphone, so these figures show how the metric behaves on hard audio, not how a dictation app will perform today. For a broader guide to human and machine accuracy, see human vs AI transcription, and for what counts as a good score, the question on what counts as a good word error rate.

WER weighs every word equally, so one wrong name can matter more than ten dropped filler words

WER treats every error as one error. A dropped "um" and a wrong surname each add one to the numerator, so the score cannot tell you which mistakes change the meaning. Kim and colleagues write that WER takes into consideration only literal correctness instead of semantic correctness, and that this matters for downstream tasks such as intent recognition and slot filling ²². Their proposed semantic distance measures meaning instead of literal matches, and the authors tested it on those downstream tasks ²².

The score also does not always track what viewers or readers think. In a 2017 study of automatic captions, word error rate correlated only 0.11 with the study's usability ratings ²³. A low WER therefore does not guarantee readable text, and a wrong drug name or client name can matter more than several dropped fillers. If particular words matter to you, count them separately: keep a list of names, numbers and terms, and check how many of them the output gets right.

Report relative WER reduction next to the absolute change, not instead of it

Papers often report improvements as relative reductions, which can sound much larger than the change in the number itself. A personalized recognizer in a 2019 Google study gave a 62 percent relative reduction in WER on speech from people with ALS ²⁴. A relative reduction divides the drop by the starting error rate. If a system falls from 10 percent to 5 percent, the absolute drop is 5 percentage points, and the relative drop is 50 percent. Both statements are true. Only the absolute change tells you how many more words are right out of every 100.

When you read an improvement, look for the starting WER, the new WER and the test set. A relative figure without them tells you little. When you write one, give the two numbers side by side.

Character error rate, semantic metrics and diarization-aware variants cover WER's blind spots

WER is one measure among several, and each variant answers a different question. The table compares the main ones.

The multi-speaker row needs more care. When a system outputs several speakers' word sequences, a plain WER does not say whether the words were attributed to the right person. Von Neumann and colleagues proposed a general framework for computing WER in this setting. If your transcripts use speaker diarization to label who spoke, score the labels and the words separately as well as together.

The comparison of WER and character error rate covers the character case in more detail. For most dictation and transcription work, report WER first, with a documented normalizer, and add a variant only when your question needs it.

Comparing two systems fairly takes the same audio, the same normalizer and the same reference

A WER comparison is only as fair as its setup. Two systems scored on different audio, with different normalizers or against different references are not comparable, even if the numbers look close. The Open ASR Leaderboard makes the setup explicit: it applies one normalization to the transcripts it scores. On its English short-form track, the average WER was 5.63 percent for NVIDIA's Canary Qwen 2.5B ²⁷, 6.05 percent for Parakeet TDT 0.6B v2 ²⁸ and 7.44 percent for Whisper Large v3 ²⁹.

Those are averages over several test sets, so they show how the systems rank on that mix of audio, not how they will do on your recordings. Leaderboards also change, and the speech recognition benchmarks guide explains how to read them. A single short test can mislead as well. The author of a 2025 9to5Mac comparison, which tested one episode, wrote that the models were all evaluated the same way but that the exact numbers would differ under slightly different methods ³⁰.

To run a fair comparison of your own, follow these steps:

1. Use the same audio for every system, recorded in the conditions you care about.
2. Make one reference transcript by hand, following written rules for fillers, numbers, names and punctuation.
3. Apply the same normalizer to the reference and to every output, and write its rules down.
4. Report the tool, its version and the number of reference words, so others can check the denominator.
5. Treat a small gap with suspicion: a short test set gives a noisy score, so repeat the test on more audio before you choose.

Read every WER number with its conditions attached: the audio, the speakers, the normalizer and the reference. If you want to check a recognizer on your own files, the free audio to text tool in the browser is one way to get a first transcript to score. For the wider question of how accurate speech recognition is in practice, the guide to speech recognition accuracy covers the real-world numbers, the guide to Whisper covers that model's measured accuracy and limits, and the guide to speech-to-text APIs compares the main providers.