Word error rate (WER) is the most widely used measure of speech recognition accuracy. It counts the words a system substituted, dropped, or added in its transcript, and divides that count by the number of words actually spoken.
Lower is better. A WER of 0% means the transcript matches what the caller said word for word. A WER of 20% means about one error for every five spoken words. For a voice agent, WER describes how well it hears, before it can understand the request or act on it.
How word error rate is calculated
The formula is WER = (S + D + I) / N.
Substitutions (S). A spoken word appears as a different word, such as "fifth" for "fifteenth".
Deletions (D). A spoken word is missing from the transcript.
Insertions (I). The transcript contains a word nobody said.
Reference words (N). The number of words in the reference transcript, a version of the same audio checked by a person. Scoring tools align the two transcripts first and then count the differences. Microsoft's speech documentation points to the sclite tool in NIST's Scoring Toolkit for this.
For example, a caller says "I want to move my payment to the fifteenth", which is nine words. The system writes "I want to move my payment to the fifth please". That is one substitution and one insertion, so WER is 2 ÷ 9, or about 22%.
Both transcripts are usually normalized before scoring, so that "15th" against "fifteenth", or a capital letter, does not count as an error. Two vendors that normalize differently can report different WERs for the same audio.
Why WER alone misleads on phone calls
Every word counts the same. In the example above, the one substitution moved a payment date. Another transcript could score the same 22% by dropping "to" and "the", and still book the right date.
Phone audio is harder. Microsoft's documentation describes contact center audio as low quality: 8 kHz, sometimes compressed, sometimes with two people on one channel. It rates contact center recognition as medium quality, under 30% WER. Voice assistants and dictation can reach under 10%.
Background speech becomes insertions. The same page links insertion errors to noisy rooms and crosstalk, and deletions to a weak audio signal. A bystander's words in the transcript matter more for a voice agent than for a caption, because the agent may act on them.
Published figures come from someone else's audio. A WER on a benchmark or in a sales deck was measured on a particular test set. Clean, read speech usually scores better than real calls from cars, shops, and speakerphones.
What to measure next to WER
Entity accuracy. The share of account numbers, dates, amounts, and names captured exactly. These are the words that change what the agent does.
Error type. Microsoft's page ties each error type to a likely cause: substitutions to missing domain vocabulary, deletions to audio level, insertions to noise. The split shows which fix to try first.
WER by condition. Separate scores for handsets, speakerphones, car kits, background speech, accents, and languages show where recognition breaks down.
Outcome. Task completion and transfers on the same calls. An agent that reads back a date before acting ("Did you say the 15th?") can recover from an error that WER still counts.
Frequently asked questions
What is a good word error rate?
It depends on the audio. Microsoft's Azure documentation treats 5 to 10% as good quality, 20% as acceptable, and 30% or more as poor. The same page expects higher rates on contact center audio, so a fair comparison runs every system on the same recorded calls.
Can word error rate be higher than 100%?
Yes. Insertions count as errors but do not add to the number of spoken words. A system that adds many words nobody said can make more errors than there were words, which pushes WER above 100%.
Why is word error rate higher on phone calls?
Phone audio carries less detail than a headset recording, and many calls come from noisy places. Narrowband audio, compression, background voices, and speakerphones all add errors. Product and plan names add substitutions when the recognizer has not seen them before.
How do you measure WER for your own contact center?
Take a sample of recorded calls that matches your real traffic, within your recording and data protection rules. People then transcribe it as the reference. Microsoft suggests 30 minutes to 5 hours of representative audio for a test. Score every system with the same tool and normalization, and report results by call condition as well as overall.
How we approach this at Zowie
Zowie's AI agents work on voice, chat, and email, with over 100 million conversations a year across our clients. Recognition is the first step of a call. After it, the language model speaks. A separate, deterministic Decision Engine returns every amount, rate, offer, disclosure, and escalation from your rules, with an audit log. To try it on your own call types, you are welcome to book a demo.
Explore: Speech analytics, Natural language understanding, Barge-in, Turn-taking, AI agent evaluation, Voice AI