Diligence & quality
WER vs CER: calculate and interpret speech recognition errors
Compare model transcripts with verified references using word error rate and character error rate. Record normalization settings so results can be reproduced and compared fairly.
- Topic
- Diligence & quality
- For
- Buyers and data owners
- Reading time
- 5 min
- Updated
Compare model transcripts with verified references using word error rate and character error rate. Record normalization settings so results can be reproduced and compared fairly.
Use the linked free tool to work through the practical steps. Keep the original files and documentation unchanged, and save your report with the date and scope of the review.
WER measures word edits against a reference
Word error rate is the number of substitutions, deletions and insertions divided by the number of reference words. If the reference is turn left and the hypothesis is turn right now, one substitution and one insertion produce two errors over two reference words: 100% WER.
The reference is the expected transcript and the hypothesis is the system output. Reversing them changes the denominator and can change the result. Use a reviewed reference; a calculator cannot determine whether the reference itself matches the recording.
CER measures character edits
Character error rate applies the same edit-distance idea to characters. Fiund uses Unicode code points after NFC normalization. Its default excludes spaces from CER; that option is visible and included in the exported report. Other implementations can choose different character or whitespace conventions.
WER uses whitespace-separated tokens in this tool. That is not a language-aware word segmenter. For languages written without consistent spaces between words, CER can provide a more interpretable starting point, but the evaluation still needs a language-appropriate reference and normalization policy.
Normalization changes the question being measured
Ignoring capitalization and punctuation may be appropriate when scoring lexical transcription. It can hide errors if punctuation is required for the downstream task. Decide how to handle casing, apostrophes, numerals, abbreviations and disfluencies before comparing systems.
Fiund offers case and Unicode-punctuation controls, always applies NFC and collapses whitespace, and records these choices in its JSON report. It does not expand numerals or apply language-specific spelling rules. An apostrophe removed by punctuation normalization changes a token; inspect examples before relying on aggregate results.
Aggregate counts, then calculate the rate
Corpus WER sums edits and reference words across utterances before dividing. Averaging utterance percentages gives short and long utterances equal weight and answers a different question. The workbench reports corpus rates from the summed counts.
Rates can exceed 100% when insertions are numerous. An empty reference has no denominator, so this tool leaves its individual rate undefined while retaining insertion counts. If the corpus includes non-empty references, those insertions still contribute to its total. Alignment ties can produce different edit breakdowns with the same minimum distance; Fiund uses a diagonal-first tie break.
Sources and further reading
Try the workflow on your own data
Open free tool →Related resources
Related reading
Open the free tool
Browser tools
Free dataset tools
Inspect local media, document dataset scope or check an evaluation sample.
Related reading
Dataset quality and delivery
5 min read
How to inventory an audio or video dataset
Build a file-level inventory before estimating volume, requesting a quote, or preparing a dataset for review. Count known durations and record the technical properties buyers need.
5 min read
Recorded hours vs usable dataset hours: how to count media volume
File-hours, unique activity-hours, speech-hours and accepted hours answer different questions. Define the unit before comparing collections or recording offers.
5 min read
How to write a dataset card buyers can evaluate
Document the dataset’s purpose, origin, collection process, limitations and permissions evidence. A useful card makes unknowns visible instead of hiding them behind marketing language.