Word error rate measures how many word-level edits separate a speech-recognition transcript from a trusted reference. A lower score means the words match more closely. It does not mean every important fact survived, nor does it tell you how much work remains before the transcript is useful.
That makes WER valuable, but easy to misuse. It is a clean way to compare systems under the same conditions. It is a poor shortcut for declaring one system universally better, or for deciding that a transcript can be trusted without review.
What is word error rate?
Word error rate, usually shortened to WER, is the standard starting metric for measuring automatic speech recognition accuracy. It compares a machine transcript with a human-checked reference transcript, often called the ground truth.
The comparison counts three types of error:
- Substitution: the transcript contains the wrong word.
- Deletion: a word from the reference is missing.
- Insertion: the transcript adds a word that was not in the reference.
Google describes WER as the industry-standard comparison method for speech recognition and also recommends considering multiple measures when the broader system matters. source
How is word error rate calculated?
The formula is:
WER = (substitutions + deletions + insertions) / words in the reference transcript × 100
Imagine a reference containing 100 words. If the machine transcript needs four substitutions, two deletions, and one insertion, the calculation is:
(4 + 2 + 1) / 100 × 100 = 7% WER
Lower is better because fewer edits are needed to match the reference.
WER can also exceed 100%. This happens when the system inserts enough extra words for the total error count to become larger than the number of words in the reference. Google explicitly notes this possibility in its accuracy documentation. source
Why normalization changes the score
Two people can calculate different WER values from the same pair of transcripts if they prepare the text differently.
Should Twenty-five and 25 count as a match? What about capitalization, punctuation, filler words, abbreviations, or accented characters? There is no honest comparison until those choices are fixed.
A basic normalization policy may:
- convert text to lowercase;
- remove punctuation;
- standardize spacing;
- decide how numbers and abbreviations are written;
- decide whether filler words count;
- apply the same rules to the reference and every machine transcript.
The exact policy depends on the task. Consistency is the important part. Changing the rules for one contender can move the result without improving a single transcript.
The reference itself also needs scrutiny. A misspelled name or missing word in the ground truth punishes a model for producing the correct text. Google recommends a carefully checked human transcript and consistent conventions before measuring accuracy. source
What is a good WER?
There is no universal number.
A useful score depends on the language, microphone, background noise, speaker, vocabulary, recording style, and what the transcript will be used for. A clean audiobook, a hurried voice note, a meeting with overlapping speakers, and technical dictation are different tests.
The honest question is not, “Is 8% WER good?” It is, “Is 8% better than the alternatives on representative audio, and are the remaining errors acceptable for this job?”
For casual notes, an error that is obvious during a quick read may be cheap to repair. For a customer name, product code, measurement, quotation, or negation, one wrong word may be the only error that matters.
This is why a speech recognition benchmark should report the test conditions beside the score. A number without the audio, language, reference policy, model version, and sample size cannot tell you much about your own workflow.
What word error rate misses
WER counts how many word edits occurred. It does not understand why one edit matters more than another.
Consider these two hypothetical errors:
| Reference | Transcript | Practical effect |
|---|---|---|
Send the draft on Friday morning | Send draft Friday morning | The sentence remains easy to understand. |
Do not send the draft on Friday morning | Do send the draft on Friday morning | One changed word reverses the instruction. |
Both examples can produce a small edit count. Their consequences are completely different.
Apple's speech research makes a related point from the reader's perspective: transcripts with similar or even higher WER can differ sharply in readability because the type of error matters. Its HEWER research treats meaning changes, readability damage, and misspelled proper nouns differently from harmless variations. source
WER also hides distribution. An average across many recordings may look respectable while one recording fails badly. If that recording represents an accent, microphone, language, or vocabulary common in your real work, the average is not reassuring.
Add a practical error review
Keep WER as the first layer, then inspect the errors that affect the intended use.
| Evaluation layer | What it tells you |
|---|---|
| Word error rate | How far the transcript is from the reference at the word level. |
| File-level scores | Whether the average hides a particularly weak recording. |
| Error categories | Whether names, numbers, negations, jargon, or required actions changed. |
| Correction time | How much work a person needs to reach a usable result. |
| Task review | Whether the transcript is safe and useful for its actual purpose. |
Microsoft's speech evaluation tools show the recognition result beside the human transcript and let evaluators inspect substitutions, deletions, and insertions. That is the useful habit to copy even when you use a different tool: do not stop at the percentage. Read the differences. source
How to run a fair ASR benchmark
An ASR benchmark becomes useful when someone else can understand exactly what was compared.
Sample size matters too. Google's current guidance recommends at least 30 minutes of representative audio and says 30 minutes to three hours is a useful range for measuring accuracy. Microsoft similarly asks for representative audio paired with human-labeled transcripts. source source
The right size still depends on the claim. A small test can expose an interesting failure. It cannot support a universal leaderboard.
Compare the workflow, not only the engine
Speech recognition is only the first part of usable dictation.
A transcript can have a low WER and still require several minutes of paragraphing, punctuation repair, name correction, and movement into the application where the work belongs. Another route may make a few more harmless word errors but require less total cleanup.
For everyday writing, track the full cost:
- time to capture the speech;
- time to receive the transcript;
- time to correct important errors;
- time to structure and format the text;
- time to put the result where it is needed.
MachinesFluent is designed around that complete Windows writing route. You can dictate in the application where you are already working, keep recurring vocabulary available, and turn rough speech into a more usable draft. The honest test is your own text: names, terminology, sentence length, preferred applications, and correction time.
For a broader product comparison, use the Windows dictation software buyer guide. If privacy changes which speech route you can use, read Local AI Models Change the Privacy Risk Profile.
Word error rate FAQ
Is a lower WER always better?
It means the transcript is closer to the reference under the chosen normalization rules. That is useful, but it does not reveal the seriousness of each error or the effort needed to make the result usable.
Can word error rate be above 100%?
Yes. A large number of inserted words can make the error count greater than the number of words in the reference transcript.
Does punctuation affect WER?
Often it does not, because punctuation and capitalization are commonly removed during normalization. The exact answer depends on the scoring policy, which should be documented and applied consistently.
What is the difference between WER and accuracy?
WER is an error rate, so lower is better. Some tools express a simplified word accuracy as 1 - WER, but that number can become misleading when insertions push WER above 100%. Reporting WER and the underlying error counts is clearer.
What is CER?
Character error rate uses character-level edits instead of word-level edits. It can be useful for languages or tasks where tokenization and spelling detail make word boundaries less informative. Like WER, it still does not judge the business importance of an error.
How much audio should an ASR benchmark use?
Use enough representative audio to support the claim you want to make. Google recommends at least 30 minutes for an accuracy measurement and suggests 30 minutes to three hours. A product decision may need more coverage across speakers, devices, accents, and environments.
Can I compare WER numbers from two different websites?
Only with great caution. Different audio, references, normalization, model versions, and test conditions can make the percentages incomparable. The fairest comparison runs both systems on the same locked test set.
Judge the errors you would actually have to fix
Word error rate is a useful ruler. It tells you how far a transcript sits from a trusted reference. It cannot tell you whether the wrong words were harmless, expensive, or dangerous.
Use WER to build a shortlist. Use real recordings and practical error review to make the decision. Then measure the complete route from speaking to finished text, because the best score is not automatically the best working experience.
Try MachinesFluent for Windows with the vocabulary and applications you use every day. Compare correction time and finished output, not one perfect demo sentence.
