Word error rate
In one sentence
Word error rate (WER) is the standard measure of transcription accuracy: the words a speech recognition system got wrong, counted as a percentage of the words that were actually said, where lower is better.
Not to be confused with Mean opinion score.
Definition
Word error rate is the usual score for transcription accuracy. It expresses the words a recognition system got wrong as a share of the words that were genuinely spoken.
Lower is better. It is also the number most likely to be quoted at you without the conditions that produced it.
To compute it you align the transcript against a reference, tally three kinds of error, and divide by how many words the reference contains.
The three error types
- Substitution. A wrong word replaced a correct one.
- Deletion. A word was said and never transcribed.
- Insertion. A word appears in the transcript that was never said.
The formula, in plain terms
- Substitutions plus deletions plus insertions, divided by the total number of words in the reference.
- Because insertions are counted, WER can go above 100 percent. A system that hallucinates extra text scores worse than one that transcribes nothing at all.
Why a flawed metric is still the one everybody uses
- It weights every word equally. Missing "the" costs what missing "not" costs, and missing "not" inverts the sentence. Missing a postcode breaks the task.
- It is sensitive to text normalization. Whether "twenty percent" and "20%" count as a match depends on the normalization applied, and providers normalize differently. That alone can move a reported figure by a point or two.
- It ignores semantic impact entirely. Two systems on identical WER can have wildly different task success rates.
Metrics that sit next to it
- CER, character error rate. Used when word boundaries are unclear, or for languages that do not separate words with spaces.
- Entity error rate, also called keyword error rate. It scores only the words that carry the task: names, numbers, product terms. Much better at predicting whether the agent works.
- Task completion rate, the metric that actually tracks the commercial outcome.
Common misconception
That a low WER means the thing will work. WER scores transcription fidelity, and fidelity is not usefulness. Take two systems: one sits at 5 percent and misses every product name, the other sits at 9 percent and gets them all. The second is the more useful of the two, because its mistakes land on words that do not change what comes back from retrieval.
Why it matters commercially
WER turns up in every procurement pack and every vendor comparison, so you have to understand it well enough to push back on it. In a vendor conversation the useful question is not what number they scored. It is which test set produced it, what normalization was applied on the way, and what the error rate looks like on the terms specific to your business.
In voice specifically
A text chat has no transcription step, so it has no WER and no error floor beneath the retrieval. Speech puts one there. That is why a voice product can be beaten by a text one running the same model over the same documents, without a single thing changing in the model or the documents.
Visual
What word error rate does not tell you
WER weights every word equally. Your customer does not.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Whisper Large-v3 records around 2.7 percent word error rate on LibriSpeech test-clean, and roughly 8 to 12 percent on real-world English audio.
2.7% clean, 8 to 12% real worldindependentWhisper Large-v3 on LibriSpeech test-clean, against reported real-world English performance, 2026 - Use it as the reference point for what a benchmark number is worth. The clean figure is a public test-set result; the real-world band is reported rather than measured once.
Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate under production conditions.
5.26% to 6.84%vendor claimDeepgram Nova-3 reported figures, 2026 - Published by the provider. A production-conditions figure is more useful than a clean-benchmark one and still not comparable to another vendor number unless the test set and the normalization match.
Human transcription, not zero, is what the practical floor is usually taken to be, since two human transcribers will disagree with each other on disfluent or noisy audio.
Not zeroindustry rangeStandard practice in speech evaluation, 2026 - Worth saying out loud in a procurement conversation. An accuracy commitment written against a zero-error reference is written against something nobody achieves.
Coval and Gradium found in 2026 that no provider led on speed and accuracy together.
No provider leads on bothindependentCoval and Gradium benchmark, 2026 - Which means a WER comparison that ignores latency is incomplete. Ask for both numbers from the same test run.
No standard for text normalization has been widely adopted in published word error rate figures, which is the main reason two vendor numbers cannot be lined up side by side.
-no reliable figureA genuine methodological gap rather than a rhetorical one, and worth stating because it explains how two providers can both honestly claim to be the most accurate.
Examples
In practice
Three providers get scored on a public benchmark and the one with the lowest WER wins the contract. Live, fewer tasks complete than the runner-up would have managed. The team re-runs the test on 200 of their own call recordings and scores entity errors separately, and the order flips: the provider they turned down handles their domain vocabulary far better. The only benchmark worth anything is your own audio.
The everyday version
Word error rate is a marks-out-of-a-hundred score for how well the system takes down what people say. Ninety-five out of a hundred sounds terrific, right up until you find that the five it drops include your company name on every single call. Ask which words go wrong, not just how many.
Usage
Who says it
- Speech engineers and machine learning practitioners, precisely, with the methodology attached.
- Procurement and vendor management, as a number to compare, and usually with no questions asked about the conditions behind it.
- Vendors, selectively, choosing the benchmark that flatters them.
Where it turns up
- Alongside accuracy commitments, the definition of the test set, acceptance criteria and occasionally penalties written into the service level.
- A well-drafted RFP requires accuracy to be demonstrated on audio the buyer supplies. A poorly drafted one accepts a published figure.
Common misuse
- Comparing WER across providers without matching test sets and normalization. The most frequent error in the category.
- Leaning on WER as the only accuracy test when entity error rate predicts task success more reliably.
- Assuming WER improvements translate into product improvements one for one. Below a threshold, further gains stop being noticeable.
Questions people ask
How is word error rate calculated?
Align the transcript against a reference transcript, count the substitutions, deletions and insertions, then divide by the total number of words in the reference. A substitution is a wrong word in place of a right one, a deletion is a word that was said and never written down, and an insertion is a word in the transcript that nobody said.
Can word error rate be more than 100 percent?
Yes. Insertions are counted in the numerator while the denominator is only the number of words in the reference, so a system that adds text can score above 100 percent. In practice that means a model which hallucinates extra words scores worse than one which transcribes nothing at all.
What is a good word error rate?
It depends entirely on the audio. Whisper Large-v3 records around 2.7 percent on the clean LibriSpeech benchmark and roughly 8 to 12 percent on real-world English, and Deepgram Nova-3 reports around 5.26 to 6.84 percent in production conditions. Human transcription is treated as the practical floor rather than zero, because human transcribers disagree with each other on disfluent or noisy audio.
Why do two vendors both claim the lowest WER?
Because there is no widely adopted standard for text normalization in reported figures, and because each vendor picks its own test set. Whether "twenty percent" and "20%" count as a match can move a published number by a point or two. Both claims can be honest and still not be comparable, which is why accuracy should be demonstrated on audio you supply.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.