Word Error Rate (WER) Calculator

Word Error Rate is the standard accuracy metric for automatic speech recognition (ASR) and, increasingly, for LLM transcription. It answers a single question: of the words a system should have produced, what fraction did it get wrong? The definition and the alignment convention this tool follows come from the NIST Speech Recognition Scoring Toolkit (SCTK, thesclite scorer used in NIST ASR evaluations).

The computation is four deterministic steps:

1. Normalise and tokenise. Both transcripts are lowercased (unless you keep case), stripped of a fixed punctuation set (unless you keep punctuation), and split on whitespace into words. In Character mode the same text is split into characters for CER.
2. Align by minimum edit distance. A Levenshtein dynamic programming matrix is built over the two token streams with unit costs: D[i][j] = min(D[i-1][j-1] + (ref≠hyp), D[i-1][j] + 1, D[i][j-1] + 1). A match costs 0; a substitution, deletion, or insertion costs 1.
3. Backtrace and count. Tracing the cheapest path back from the bottom-right corner recovers each operation, giving the number of substitutions (S), deletions (D), insertions (I), and correct matches (C). This is the SCTK counting convention.
4. Compute the metrics. The headline figure is WER = (S + D + I) / N, where N is the reference length. Accuracy is 1 − WER. The tool also reports MER = (S + D + I) / (S + D + I + C), from Morris, Maier & Green (2004), which stays bounded to 100% even when WER does not.

The sum S + D + I is exactly the Levenshtein edit distance between the two token streams, so the calculator recomputes that distance a second, independent way and confirms the two agree — the “verified ” badge in the results footer. A deletion means a reference word is missing from the hypothesis; an insertion means the hypothesis added a word that isn’t in the reference. Because insertions are counted but only reference words are in the denominator, WER can rise above 100% — a real property of the metric, not an error. The colour-coded alignment shows precisely which words were substituted, dropped, or added, so you can see whether the errors cluster on proper nouns, numbers, or filler words and act accordingly (for example, adding a custom vocabulary).

• NIST — Speech Recognition Scoring Toolkit (SCTK / sclite), the reference WER implementation
• Morris, Maier & Green (2004) — From WER and RIL to MER and WIL (Interspeech)
• Manning, Raghavan & Schütze — Edit (Levenshtein) distance, Introduction to Information Retrieval

The WER formula and alignment follow the NIST SCTK (sclite) WER definition; the metrics were last cross-checked against these sources on 2026-07-08. The two worked examples above (22.22% and 125%) are reproduced exactly by the calculator.