What is word error rate (WER)?
Word error rate is the standard measure of transcription accuracy for a speech recognition system, expressed as the proportion of words it gets wrong. It counts three error types, substitutions, deletions, and insertions, against a human-verified reference transcript of the same audio.
WER is reported as a percentage, and it can exceed 100 percent, because insertions add errors without adding reference words. In voice AI it is usually computed per utterance and averaged across a corpus, since one call mixes clean speech with hold music, crosstalk, and codec artifacts.
How word error rate is calculated
Calculation runs in three steps. The hypothesis transcript produced by the recogniser is aligned to the reference transcript word by word, using the edit distance that minimises total changes. Each mismatch is then classified as a substitution (one word heard as another), a deletion (a spoken word missing), or an insertion (a word that was never said). The errors are summed and divided by N, the number of words in the reference:
WER = (S + D + I) / N
A worked example: a 120-word reference transcript yields 6 substitutions, 3 deletions, and 1 insertion. That is 10 errors over 120 reference words, so WER is 8.3 percent.
The number rarely stays inside the speech stack. A mistranscribed order number pushes the call to a person and lifts escalation rate, which drags down chatbot containment rate on the voice channel and, further downstream, the queue’s resolution rate.
What counts as an error and what does not
Substitutions: One word is heard as another, the error type most likely to change meaning, as when “eight” becomes “late”.
Deletions: A reference word is absent from the hypothesis, common when speech is clipped by barge-in or a dropped packet.
Insertions: The recogniser emits a word nobody spoke, usually from background speech or breath, which is why the rate can pass 100 percent.
Normalisation: Casing, punctuation, filler words, and number formatting are settled before scoring, and changing that policy changes the score without changing the audio.
Not counted: Meaning, severity, and speaker intent carry no weight, so a missed “not” scores identically to a missed “the”.
Word error rate vs character error rate vs sentence error rate vs intent accuracy
Buyers evaluating voice agents are handed competing accuracy figures and told they describe the same thing. Character error rate scores the same alignment at the character level, which suits agglutinative languages and spelled-out identifiers. Sentence error rate scores each utterance as wholly right or wholly wrong, which is harsher and closer to how a caller experiences a failure. Intent accuracy scores whether the request was classified correctly, and it can stay high while the transcript underneath is messy. Word error rate sits beneath all three as the transport-layer measure: it reports what the recogniser heard before any meaning was attached.
What it counts | What it misses | Typical benchmark | |
|---|---|---|---|
Word error rate | Substitutions, deletions, insertions per reference word | Which errors actually change the outcome | Low single digits on clean read speech, higher on telephony |
Character error rate | Character-level edits between hypothesis and reference | Word-boundary failures and meaning loss | Usually a fraction of WER on the same audio |
Sentence error rate | Utterances containing any error at all | Severity and count of errors within an utterance | Well above WER on the same corpus |
Intent accuracy | Whether the request was routed to the right intent | Verbatim fidelity, names, numbers, addresses | Reported per intent set, so it travels badly between systems |
If you are diagnosing the recogniser itself, word error rate is the number to track, because it isolates the audio path from the reasoning path. If you are reporting to the business on whether callers got what they wanted, intent accuracy and containment describe that outcome more honestly.
Why word error rate matters for customer experience
When word error rate goes unmeasured, failures get blamed on the wrong layer. The agent looks like a poor reasoner when in fact it reasoned correctly over a transcript that said “card” where the caller said “cart”. Customers experience it as being asked to repeat themselves, spelling a surname letter by letter, and eventually hanging up, which surfaces as a rising call abandon rate and as repeat calls that inflate contact rate.
There is a real tradeoff. Lowering the cost of an error usually means adding confirmation turns (“I heard four two one, is that right?”), and every confirmation adds seconds to a call the customer already thinks is slow. A model tuned on the dominant accent in your audio also reads as accurate in aggregate while failing a minority of callers consistently, and the average hides them.
How is word error rate benchmarked?
Benchmarking requires a fixed corpus and a fixed reference transcript, or the figure describes the test set more than the system. The public reference point is the LibriSpeech ASR corpus, roughly a thousand hours of read English audiobook speech split into clean and noisier test sets, where strong modern systems report approximately 2 to 6 percent WER across the two splits.
Production measurement is a different exercise. Sample real calls across accents, handset types, and background conditions, have humans transcribe them verbatim under one normalisation policy, then score by segment as well as in aggregate. Expect telephony audio, narrowband and compressed, to score several points worse than any published read-speech figure quoted at you.
How AI agents change word error rate
Two mechanisms have changed what the number means. End-to-end models fold acoustic and language modelling into a single network trained on far more audio, so domain adaptation now happens through biasing lists, hotword hints, and fine-tuning, and the hand-built pronunciation lexicon has largely disappeared. The second is downstream repair: a language model reading the transcript reconstructs “reset my password” from a garbled version using conversational context, so scattered errors no longer guarantee a wrong action.
The consequence is that WER became a leading indicator of trouble rather than the outcome itself. Teams read it beside the conversation-level numbers that platforms tracking containment rate report, using WER to explain why an outcome moved and the outcome metric to decide whether it mattered.
How to reduce word error rate
Start with audio capture, because no model recovers information the codec discarded: wideband where the carrier allows it, echo cancellation, and a hard look at the IVR prompts that encourage callers to talk over the agent. Then supply domain vocabulary. Product names, SKUs, drug names, and city names are the words a general model has never seen, and biasing lists move those errors more than a model upgrade does. Third, fix normalisation once and freeze it, so month-over-month comparisons mean something.
Governance sets the boundary on all of it. GDPR treats a call recording as personal data, so consent capture and retention windows decide how long your evaluation set is even allowed to exist, and human transcription is the constraint that bites hardest: verbatim reference transcripts take several hours of labour per hour of audio, which keeps evaluation sets small and lets them go stale as the product changes.
Word error rate and voice self-service metrics
Word error rate is an input to the voice channel’s economics. Every utterance the recogniser mangles is an answer the agent cannot ground, which is why voice deflection rate tends to move within a quarter of a serious transcription fix and barely responds to prompt tuning.
The effect compounds over time. Callers who are misheard twice stop trusting the line and route around it, and in subscription businesses that pattern shows up months later in customer churn rate, long after the transcription defect that started it was forgotten.
What does word error rate mean in plain terms?
Word error rate, almost always shortened to WER, stands for what it sounds like: the rate at which words come out wrong. Think of it as a dictation test scored by a strict marker who compares the machine’s typed page against what was actually said and counts every wrong word, every missing word, and every word invented.
Without that score, you are guessing. An agent that cheerfully cancels the wrong order looks like a broken policy engine until someone reads the transcript and finds the account number was wrong by one digit.
The tradeoff is that the strict marker has no judgment. A missed “please” and a missed “no” both cost one point, so a system can improve its score while getting worse at the sentences that decide whether a customer is charged.
Common word error rate mistakes
Quoting a published figure as if it were yours is the first. Read-speech corpora are recorded in quiet rooms at full bandwidth, and a compressed mobile call in a car shares almost none of those conditions, so the two numbers are not measuring the same difficulty.
Averaging across utterance lengths is the second. WER divides by reference words, so a short utterance has a tiny denominator: one error in a five-word confirmation is 20 percent, and a handful of those will dominate a corpus average built from long turns.
Weighting every word equally in your analysis is the third. The errors that cost money cluster in names, numbers, dates, and negations, so a per-entity error rate on those fields tells you more about risk than the headline figure ever will.
The fourth is optimising the metric itself. Chasing a lower score with normalisation tricks improves the report while callers see no change, the same trap described in these trust metrics for AI support.
Frequently Asked Questions
What is a good word error rate for a voice agent?
A good word error rate depends entirely on the audio conditions. On clean, wideband speech, single-digit percentages are ordinary for current systems. On compressed telephony with accents, background noise, and spelled-out identifiers, the same system will score meaningfully worse. Judge any figure by the corpus it came from, and always re-measure on your own recorded calls.
What is the difference between word error rate and character error rate?
Word error rate scores whole-word mismatches after aligning two transcripts, while character error rate scores the individual character edits needed to turn one string into the other. Character error rate is gentler on near-misses and better suited to languages with long compound words, spelled identifiers, and rich morphology. Both use the same alignment logic at different granularity.
Word error rate vs accuracy: are they the same thing?
Word error rate and accuracy are complements only in the simplest case. People often quote accuracy as 100 minus WER, which breaks down because insertions can push WER above 100 percent, leaving negative accuracy. Accuracy also gets used loosely to mean intent classification quality, a conversation-level measure that says nothing about verbatim transcription.
Can word error rate be higher than 100 percent?
Word error rate can exceed 100 percent, and this surprises people. The denominator counts only the reference words, while the numerator counts substitutions, deletions, and insertions. A recogniser that hallucinates extra words from background noise can generate more errors than there were words spoken, which produces a score above 100.
How do you calculate word error rate manually?
Calculating word error rate manually means aligning the machine transcript against a verbatim human transcript, marking each substitution, deletion, and insertion, adding them together, then dividing by the number of words in the human transcript. Multiply by 100 for a percentage. Decide your punctuation, casing, and number formatting rules before you start scoring.
Does a low word error rate mean the AI agent understood the caller?
A low word error rate means the words were transcribed correctly, which is necessary but insufficient. The agent still has to classify the intent, retrieve the right policy, and take the correct action, and any of those can fail on a perfect transcript. Track transcription quality and task completion separately, because they fail for different reasons.