Word error rate (WER) The standard score for speech recognition, and what it hides.
Short answer. Word error rate (WER) is the standard score for speech recognition: the number of word substitutions, deletions and insertions needed to turn the system’s transcript into the correct one, divided by the number of words in the correct transcript. A WER of 10% means about one word in ten is wrong. Because insertions count, WER can exceed 100%. It treats every word equally, so a wrong “Tuesday” costs the same as a wrong “the”, and it depends on how text is normalised. Jurafsky and Martin put today’s best systems near 2% on clean English read speech and 5.8 to 11% on meetings and phone conversations.
01
The formula, with a worked example
WER equals 100 times (insertions + substitutions + deletions) divided by the number of words in the correct transcript. The textbook notes that the standard tool is NIST’s free sclite script.
Example. Reference: “send the revised draft to Mr Rao by Friday” (9 words). Hypothesis: “send revised draft to Mister Rao by Friday morning”. The system dropped “the” (one deletion), wrote “Mister” for “Mr” (one substitution) and added “morning” (one insertion). Three errors over 9 words is 33.3%.
02
What WER hides
OpenAI’s Whisper paper says WER penalises innocuous differences in transcript style, so the team built a text normaliser before scoring. The textbook lists normalisation rules such as removing filled pauses, expanding contractions and spelling out numbers. For mixed Hindi and English, the MUCS 2021 challenge added a transliterated WER that counts an English word as correct whether it appears in Latin or in its native-script form.
Issue
Example
Effect
Every word counts the same
A wrong “Tuesday” and a wrong “the”
A harmless slip and a wrong date score alike
Normalisation
“Mr” and “Mister”, “colour” and “color”, “100” and “one hundred”
Style differences are counted as errors unless text is normalised first
Script
Hinglish “kal” and “कल” for the same word
Exact matching counts a correct word as wrong
Insertions
The system adds words in silence
WER can pass 100%
Names and amounts
A wrong figure in a rupee amount
One word error, possibly the one that matters most
03
Typical scores
The scores below are word error rates. Languages written without spaces, such as Mandarin, are often scored by character error rate instead.
Test
Word error rate
Source
English audiobooks (LibriSpeech clean)
About 2%
Jurafsky and Martin
AMI meetings, Switchboard and CALLHOME conversations
MUCS 2021 baseline, Hindi-English and Bengali-English code-switching
32.45%
MUCS 2021 organisers
04
How to test a tool on your own meetings
Take ten minutes of a real meeting, with the usual speakers and room.
Have a person write the correct transcript, including how you want Hinglish written.
Run each tool on the same audio and compute WER with a script such as sclite or an open-source library.
Count separately the errors in names, numbers and dates, which matter more than the total.
Repeat for an accent or a speaker the tool finds hard, and for a noisy stretch.
For scale, assume an hour has 9,000 spoken words, an assumption for the arithmetic: 10% WER is then about 900 word errors, so the transcript needs a human pass either way.
05
Where ANVE stands
ANVE does not rely on a transcript score alone for minutes. Every action item must quote the transcript line it came from, and a person confirms each owner before the item is approved, so a transcript error can be caught where it matters.
06
Questions people ask
What is a good WER?
There is no universal threshold. Clean English read speech is near 2% for the best systems and meetings run 5.8 to 11%, so judge by what an error costs you and test your own audio.
Can WER be over 100%?
Yes. Insertions count as errors, so a transcript with many extra words can score above 100%. Whisper large-v2 scored 102.7% on Gujarati in OpenAI’s paper, which means it was effectively failing.
Is a lower WER always better?
Not always. WER weights every word equally, depends on normalisation, and can penalise correct words written in a different script or style. Look at errors in names, numbers and dates as well.
How do you score Hinglish?
The MUCS 2021 challenge used a transliterated WER that accepts an English word in either Latin or native script. If you score your own tests, decide the same rule in advance.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
