AI Transcription Accuracy: What WER Really Means

If you have ever seen a speech-to-text vendor claim "95% accuracy," they are almost always referring to Word Error Rate (WER) — the dominant way the industry measures how well a transcript matches what was actually said. Understanding WER tells you what those marketing numbers really mean, why a transcript can score well yet still feel wrong, and what to test before you trust an AI notetaker with your meetings or sales calls.

What WER actually measures

Word Error Rate counts how many words a system gets wrong relative to a verified "ground truth" reference transcript. The formula is simple:

WER = (S + D + I) ÷ N

• S — Substitutions: a word swapped for the wrong one ("noose" transcribed as "moose").
• D — Deletions: a spoken word dropped entirely ("turn it around" becomes "turn around").
• I — Insertions: a word added that nobody said ("SAT" becomes "essay tea").
• N — Total words in the reference transcript.

So if someone speaks 100 words and the system makes 8 combined errors, the WER is 8%. A common rule of thumb: subtract WER from 100 to get a rough "accuracy" figure (8% WER ≈ 92% accurate). Behind the scenes, the errors are counted using sequence alignment (Levenshtein distance), which finds the minimum number of edits needed to turn the machine output into the reference — you cannot just eyeball the differences.

A quick worked example

Reference: "please send the contract by Friday afternoon" (7 words). Suppose the transcript reads "please send the contract Friday at noon." That is one deletion ("by"), one substitution ("afternoon" → "noon"), and one insertion ("at"). Three errors over seven words gives a WER of about 43% — high, because short samples are unforgiving.

What counts as a good WER score

Accuracy depends heavily on audio conditions, so context matters more than any single benchmark:

• Clean, scripted audio: top models like OpenAI's Whisper Large-v3 hit around 2.7% WER on the LibriSpeech test-clean benchmark, and Deepgram's Nova models report median WER in the 5–7% range.
• Real-world conversation: on meetings, podcasts, phone calls, and interviews, the same English models typically land somewhere around 8–12% WER.
• Hard conditions: heavy background noise, overlapping speakers, strong accents, or specialized jargon can push WER far higher — models that score above 95% on benchmarks can fall to 70% or lower in messy live audio.

As a practical guide: under 5% WER is excellent, 5–10% is good and very usable, 10–20% is readable but needs review, and above 20% becomes frustrating for skimming or analysis.

Why benchmark numbers mislead

The biggest trap is assuming a vendor's headline WER applies to your audio. Academic benchmarks like LibriSpeech use clean, read speech recorded in controlled conditions. Your sales calls have crosstalk, dial-in static, filler words, brand names, and accents none of those datasets contain. Production accuracy is routinely several points worse than the benchmark, so always test on a sample of your own recordings rather than trusting a published figure.

The limitations of WER

WER is the best single metric we have, but it is a blunt one — and knowing its blind spots is the whole point.

• It treats every error equally. Mishearing "cat" as "bat" costs the same as dropping the word "not" — even though one flips the meaning of a sentence and the other is trivial.
• It ignores punctuation and formatting. A wall of correctly transcribed words with no punctuation can score a perfect 0% WER yet be painful to read.
• It ignores speaker labels and structure. WER does not measure diarization (who said what), timestamps, or paragraph breaks — all of which determine whether a transcript is actually usable.
• It is highly sensitive to normalization. Differences in how "$5" vs "five dollars" or "Dr." vs "Doctor" are written can inflate WER even when the system heard everything correctly. Vendors using different normalization rules cannot be compared apples-to-apples.

Why it matters for call recording and analysis

For an AI notetaker or call-QA tool, transcription accuracy is the foundation everything else stands on. AI summaries, action items, and sales-call scoring against a checklist are only as reliable as the words underneath them. A high WER on the names, numbers, and objection-handling phrases that matter can quietly corrupt a coaching insight or a candidate-evaluation signal.

This is why production tools often combine engines and a fallback path rather than relying on one model. MeetGrade, for example, transcribes Zoom, Google Meet, and phone calls through Deepgram with a Groq fallback, then layers QA scoring, conversation metrics, and AI coaching on top — and because every transcript is visible alongside the audio, you can verify any score against what was actually said instead of taking a number on faith. For interview and candidate use, that evidence trail keeps the analysis grounded in competencies and structured-interview signals — it is decision-support, explicitly not lie-detection or facial-emotion reading.

How to evaluate accuracy yourself

• Test on your real audio, not vendor demos — ideally your noisiest, most accent-diverse calls.
• Read the transcript, not just the score. Check whether names, prices, and negations are correct.
• Look beyond WER: speaker separation, punctuation, and timestamp quality often matter more for real workflows.
• Use a custom vocabulary for product names and jargon where the tool supports it — it meaningfully cuts substitution errors.

WER is a useful compass, not a guarantee. Treat it as one input, verify accuracy on your own recordings, and judge a system by whether the transcript supports the decision you need to make. If you want to see transcription accuracy in the context of real call analysis and QA, you can try MeetGrade on your own meetings and check every insight against the words on the page.