What Is Word Error Rate (WER)?

Direct definition

Word error rate (WER) is the standard research metric for speech recognition. It counts substitutions, deletions, and insertions against a reference transcript, then divides by the number of words in that reference.

The sections below explain how word error rate is used in transcription and subtitle work, what people often mix up, and which VideoText workflow applies when the next step is a file or review pass.

Key takeaways

• WER = (substitutions + deletions + insertions) / reference words.
• Lower is better. WER can exceed 100% if the system inserts many extra words.
• Always name the dataset and text-normalization rules next to a WER figure.

How WER is computed

Align the hypothesis to the reference with an edit distance. Each wrong word, missing word, or extra word is an error. Divide by N, the reference word count.

Why WER matters

It lets researchers compare systems on the same audio and the same reference. The Whisper paper reports many WERs so readers can see robustness, not just LibriSpeech Clean.

The thing people often get wrong

Converting WER to “percent accurate” (100 − WER) hides insertions and treats every word as equal. A 2.5% WER on clean audiobooks does not mean 97.5% accurate captions on a Zoom call.

Frequently asked questions

Is VideoText’s 2.52% pilot the same as Whisper’s LibriSpeech score?

No. The VideoText research page reports a small Phase 1 pilot on Whisper small. The Whisper paper’s Table 2 LibriSpeech Clean figure for Large V2 is 2.7% WER — a different model and evaluation. The two numbers are not interchangeable.

Related terms

• character error rate
• transcription accuracy
• automatic speech recognition
• hallucination
• All glossary terms

Sources

1. OpenAI: Whisper paper (WER tables)
2. VideoText: AI Transcription Accuracy Benchmark — Phase 1