Updated 2026-09-16
ASR word error rate (WER) is the percentage of words incorrectly recognized by an automatic speech recognition system — combining substitutions, deletions, and insertions — compared to a reference transcript, used to benchmark speech-to-text accuracy.
Speech recognition researchers measure accuracy with word error rate: WER = (Substitutions + Deletions + Insertions) ÷ Total words in reference transcript. A ten-word sentence with one wrong word substituted scores 10% WER. Insertions count extra words the ASR hallucinated; deletions count words it missed entirely.
Lower WER is better. Published benchmarks on clean American English audiobook speech often show single-digit WER on leading models — but hiring interviews are not audiobooks. Accents, code-switching, background noise, cheap phone microphones, and domain vocabulary ('nursing council registration', 'UPI', 'pick-face zone') push real-world WER higher than lab charts suggest.
Interview engines do not expose WER directly to recruiters, but they feel it as wrong follow-up questions, missed knockout answers, or scores attached to transcript text the candidate never said. ASR quality is a first-class hiring risk, not an engineering footnote.
Models trained disproportionately on certain accent distributions perform worse on others — documented in academic and industry evaluations across Indian English, regional language-influenced English, and multilingual code-switching common in India hiring pipelines. A candidate judged on a garbled transcript is evaluated on words they did not speak.
Mitigations include accent-robust models, custom vocabulary for role terms, confidence thresholds that trigger repeat requests ('I didn't catch that — could you say it again?'), human review for low-confidence segments, and offering language options where the product supports them.
Fair hiring reviews should ask vendors for WER or accuracy evaluations on speaker populations matching your candidate pool — not only default US English test sets.
| Metric | What it captures |
|---|---|
| WER | Word-level errors vs reference transcript |
| CER | Character-level — used for some languages |
| Real-time factor | Speed of processing vs audio duration |
| Confidence scores | Model uncertainty per token — useful for gating |
| Semantic accuracy | Meaning preserved despite minor word errors — harder to automate |
There is no universal threshold — depends on question type and downstream handling. Critical structured fields need near-perfect capture; behavioural answers may tolerate higher WER if meaning stays intact.
WER differences across demographic speaker groups indicate disparate technical performance — related to fairness risk, though legal analysis requires broader context.
Review and correction workflows help for high-stakes decisions but do not scale to ten-thousand-candidate bulk screens — prevention at capture time matters more.
Products may show qualitative confidence or 'audio quality' warnings rather than raw WER — ask vendors what underlying metrics drive those indicators.
Head back to Hiring glossary or start now.
· 3 free credits · pay per interview · nothing recurring