word error rate (WER)
A metric to evaluate the accuracy of speech recognition systems, representing the percentage of words incorrectly transcribed.
Learn
When to use it
Use word error rate (WER) when you need to quantify the accuracy of a speech recognition system. WER brings together substitutions, insertions, and deletions to provide a clear metric for evaluating transcription errors, helping to compare systems or track improvements over time.
Quick example
In a speech-to-text application like Google's Speech-to-Text API, developers measure the system's transcription accuracy by calculating the WER. The API outputs a transcription, and developers compare it to a human-verified transcript to determine the percentage of words that were incorrectly transcribed. In this setup, WER is the metric that quantifies the system's accuracy.
Ecosystem
WER sits alongside other evaluation metrics and processes in speech recognition, such as precision, recall, and the alignment of audio segments.
transcription → WER → evaluate accuracy
Misconceptions
| Misconception | Rebuttal |
|---|---|
| WER only counts substitutions | WER includes substitutions, insertions, and deletions |
| Lower WER always means better quality | Lower WER may overlook context or speaker nuances |
Trade-offs
- Accuracy insight — may not capture context or speaker intent
- Comparability — can oversimplify nuanced errors in complex audio