Calculator inputs
Formula used
S means substitutions, D means deletions, and I means insertions. N represents reference characters after selected normalization rules. Lower scores usually indicate more accurate model predictions overall.
How to use this calculator
Choose single mode for one reference and prediction pair. Choose batch mode for many dataset records. Apply identical normalization rules across every model comparison.
- Enter the correct reference text.
- Enter the model-generated hypothesis text.
- Select segmentation and normalization settings.
- Enable alignment, confusion, or matrix details.
- Calculate, review, copy, or export results.
Example data
| Reference | Hypothesis | Errors | Reference length | CER |
|---|---|---|---|---|
| machine learning | machine learing | 1 deletion | 16 | 6.25% |
| hello world | hello word | 1 deletion | 11 | 9.09% |
| AI model | ai models | 1 insertion | 8 | 12.50% |
Interpretation guide
A zero CER indicates an exact normalized character match. Higher scores reveal more insertions, deletions, or substitutions. CER may exceed one hundred percent with many insertions.
Character accuracy is bounded at zero for readability. Semantic quality can still differ despite matching characters. Always inspect examples before making deployment decisions confidently.
Limitations
CER measures surface edits instead of meaning or usefulness. Normalization choices can significantly alter reported evaluation scores. Consistent preprocessing keeps model comparisons fair and reproducible.
Very long samples require substantial dynamic-programming memory. This page limits normalized samples to 1,200 tokens. Split larger documents into sensible evaluation segments first.
Frequently asked questions
What is Character Error Rate?
CER measures character-level differences between reference and predicted text. It counts substitutions, deletions, and insertions using alignment. Lower values usually indicate stronger recognition performance overall.
Can CER exceed 100 percent?
Yes, CER can exceed one hundred percent. Heavy insertions may outnumber reference characters significantly. This behavior is mathematically valid and expected sometimes.
What is the difference between CER and WER?
CER evaluates individual characters within aligned text sequences. WER evaluates complete word tokens instead. CER better exposes spelling and recognition details directly.
Should spaces count as characters?
Spaces commonly count during standard CER evaluation. Some research pipelines remove them before scoring. Use the same rule across every experiment consistently.
Which segmentation mode should I use?
Grapheme clusters suit multilingual text and combined symbols. Code points expose underlying Unicode elements separately. Bytes help inspect storage-level encoding differences during debugging.
What does Unicode normalization change?
Unicode normalization standardizes equivalent character representations before comparison. NFC preserves normal canonical text composition. NFKC also folds many compatibility characters together.
How are batch averages calculated?
Micro CER combines all errors and reference characters. Macro CER averages each sample's individual CER. Both reveal different dataset performance patterns and risks.
Does this calculator store submitted text?
This standalone file performs calculations during each request. It includes no database storage by default. Hosting logs may still follow server configuration rules.
How should models be compared fairly?
Use identical references, normalization, segmentation, and averaging methods. Report both aggregate scores and difficult examples. Document every preprocessing decision for reproducible evaluation later.