Transcript scoring tool
Word error rate, without mystery math.
Compare a reference transcript and a hypothesis transcript with plain dynamic-programming edit distance. The calculation stays in your browser and makes every substitution, deletion, and insertion visible.
Reference + hypothesis
Choose the normalization before you read the number.
The methodology is explicit because changing case or punctuation rules changes the edit count. This page computes WER on words and CER on characters using the same chosen normalization.
Inputs
Keep the transcripts local.
Methodology
Say exactly how the score was made.
This calculator follows the standard minimum-edit-distance family used in ASR evaluation. The authoritative open reference is NIST SCTK, and IraVoice publishes a separate benchmark that reports WER for Apple SpeechAnalyzer and whisper.cpp on a disclosed Mac setup.
Current scoring settings
Unicode NFKC normalization · case folded to lowercase · Unicode punctuation and symbols stripped · whitespace collapsed before scoring- Apply Unicode NFKC normalization, trim the ends, and collapse repeated whitespace to single spaces.
- Lowercase both transcripts before scoring.
- Replace Unicode punctuation and symbol characters with spaces before tokenization.
- Split WER on whitespace-delimited words.
- Compute minimum-edit alignment with dynamic programming and report substitutions, deletions, and insertions.
- Compute CER on the same normalized text after removing whitespace so spacing does not dominate the character score.
Interpretation
WER is not the whole product story.
A lower WER can matter, but the same recognition engine can still feel better or worse depending on latency, formatting, domain vocabulary, punctuation restoration, multilingual support, accessibility needs, and whether the test audio matches real use.
Use this page to compare transcript disagreement under named rules. Do not treat one WER number as a final verdict on usefulness or fairness.