What is word error rate (WER)?

WER is the number most speech recognition comparisons lead with, so it’s worth knowing what’s inside it.

You need two transcripts: the reference, which a person checked, and the hypothesis, which the recogniser wrote. Line them up word by word with the fewest possible edits. Every word that was swapped is a substitution, every missing word a deletion, every extra word an insertion. Add those up and divide by the number of words in the reference.

Lower is better, and zero is perfect. Because insertions count, a bad enough system can go above 100%.

The number moves a lot with how you score. Punctuation, capital letters, “Rs” versus “रुपये”, “10” versus “ten”: unless both transcripts are normalised the same way first, those all count as errors. That’s why serious reports say “normalised WER” and publish the normaliser.

Hindi adds its own traps. The same word can be spelled more than one way in Devanagari. English words inside Hindi speech can be written in Roman or in Devanagari, so “delivery” and “डिलीवरी” are a mismatch even when the recogniser heard it right. A fair comparison picks one convention and applies it to every system.

WER also treats every word as equal. Dropping “को” and dropping a digit from a phone number both count as one error, but only one of them breaks your call. For phone agents, check the words that carry the job, like amounts, dates, names and numbers, alongside the headline figure.

And always ask which audio it was. Read speech in a studio scores far better than an 8 kHz call from a moving bus.