Word Error Rate, usually written as WER, is one way to measure speech recognition accuracy. It is the percentage of words in a transcript that do not match what was actually said. A 2% WER means 98 of 100 words were transcribed correctly; a 5% WER means 95 of 100. Lower is better.
That is all it is: a count of errors in one specific test, run under one specific set of conditions. Marketing claims about accuracy almost always cite WER, and WER is almost always measured under conditions that bear little resemblance to actual use.
How word error rate is actually measured
WER is calculated by taking a recording of someone reading text aloud in a quiet room, running the speech-to-text engine on it, and counting the percentage of words that came out wrong. The standard test uses professional recordings, professional microphones, and text that was written in advance to be read aloud — almost always in one of a few standard varieties of English.
That is why it is useful: it is reproducible and comparable across engines. Different engines run on the same benchmark, and the percentage is an apples-to-apples measure of transcription quality under those exact conditions. This is the foundation for how on-device speech recognition and cloud services are both measured.
It is also why it is misleading in practice.
What a published WER does not measure
A speech recognition engine's actual performance in use depends on factors that WER does not include:
- Microphone quality. The benchmark uses professional recording equipment. A laptop microphone or a headset will produce different results, sometimes dramatically different.
- Background noise. Benchmarks are usually measured in quiet studios. Real dictation happens in offices with keyboards clicking, in cars, in coffee shops.
- Accent and voice characteristics. Benchmarks are usually trained on a narrow range of English-language speech, usually from North America or the UK. A speaker with a different accent, or speaking a language the model was not trained on, will see higher error rates than the published benchmark.
- Speaking style. Benchmarks use text read aloud in formal speech. Conversational speech, stream-of-consciousness thinking, technical jargon and rapid pacing all produce different error rates than a person reading a prepared script.
- The specific words you use. An engine trained to recognize finance jargon will perform worse on medical terminology, and vice versa.
Why WER compounds in real use
This is the part marketing glosses over. If an engine has a 2% error rate, that means 2 out of 100 words are wrong. But most dictation happens in sentences or paragraphs, not 100-word runs. A 50-word sentence with a 2% error rate will have roughly one error in it — meaning a sentence-level accuracy of about 63%. A 100-word paragraph will average two errors. The small percentage stacks up quickly.
When accuracy actually matters
Accuracy is real and measurable on benchmarks. Where it matters in practice is narrower than marketing suggests:
- Jargon-heavy work. An engine trained on medical terminology will handle medical transcription more accurately than a general engine.
- Noisy environments. An engine designed to filter background noise will perform better where there is noise.
- Uncommon languages. An engine trained specifically on your language and accent will perform better than one trained on a different variety.
For general English-language dictation in a reasonably quiet setting, the difference between a 2% and 3% WER published benchmark is so small that microphone quality and speaking clearly matter far more than the engine choice.
Where to find an engine's actual WER
Most companies that publish WER include a link to the paper or documentation that explains how they measured it. Read the methodology, not the number: if they measured on scripted English from native speakers and you are dictating conversational medical notes with an accent, the gap between their benchmark and your reality is probably large.