Quick answer
WER = (substitutions + deletions + insertions) ÷ words in the human reference. CER uses characters instead. Lower is better, but the score is meaningful only for the audio and normalization rules you tested.
Turn your recording into text
Drop or choose an audio or video file. Upload starts immediately; your transcript opens on TranscribeNext.
MP3, WAV, M4A, MP4, WebM
Build a trustworthy reference
Choose clips from your real workload, including ordinary noise and speakers. A human should correct each reference transcript while listening to the audio. Decide in advance whether punctuation, filler words and numerals count. Apply the same rule to both the reference and each machine output.
What the score catches
A substitution replaces a reference word; a deletion misses it; an insertion adds a word that was not said. WER counts all three and divides by the reference word count. For example, one substitution in a ten-word reference yields 10% WER. It does not follow that every future recording will have 90% accuracy.
Japanese and Korean need extra care
Word boundaries depend on the tokenizer. For Japanese and Korean, two legitimate segmentations can produce different WER values. Character error rate is often a more reproducible supplementary metric, but normalize spaces and punctuation consistently. Report both the rules and sample size.
Use the free local checker
Paste a human-corrected reference and an AI transcript into our transcript accuracy checker. The calculation runs in your browser. It reports insertions, deletions, substitutions, WER and CER; no text is sent to our server by the tool. It is a measurement aid, not a substitute for listening.