How Accurate Is Whisper? Real WER Numbers (2026)

The Short Answer: Whisper's Accuracy in Numbers

Accuracy in speech recognition is measured with word error rate (WER): the percentage of words the system gets wrong. Lower is better. Here is where Whisper stands on published benchmarks as of 2026:

These figures come from public benchmark compilations (artificialanalysis.ai, OpenAI's own evaluations, independent WER roundups) as of 2026. Whisper is also an industry reference point: MLCommons added it to MLPerf Inference v5.1 as the reference speech-recognition benchmark in September 2025.

The spread between 2.7% and 17.7% is the whole story: the model is the same, the audio is not. Knowing which scenario your use case falls into tells you how accurate Whisper will be for you.

What Word Error Rate (WER) Actually Measures

WER compares a machine transcript against a human-verified reference transcript and counts three kinds of mistakes:

• Substitutions - the wrong word appears ("their" instead of "there")
• Deletions - a spoken word is missing from the transcript
• Insertions - a word appears that was never spoken

Add the three up, divide by the words in the reference, and you get WER: 5% means 5 mistakes per 100 words. Two caveats before you compare numbers across articles:

• WER treats all errors equally. Transcribing "can" as "can't" counts the same as a harmless "uh" insertion, even though one flips the meaning of a sentence.
• WER depends on the test set. A model can score 3% on audiobook-style audio and 15% on the same speaker in a noisy cafe. Any accuracy claim without a description of the audio is close to meaningless.

That is why the honest answer to "is Whisper accurate" is a range, not a single number.

Whisper vs Human Transcribers

The most striking benchmark result: on clean audio, Whisper is now more accurate than professional human transcribers. Humans score around 4-6.8% WER on the standard benchmarks where Whisper large-v3 scores about 2.7%. Humans mishear words, lose focus, and make typos; a model does not get tired.

That said, humans keep two real advantages:

• Context and judgment. A human who knows the subject matter will correctly render a niche product name or an inside joke that Whisper renders phonetically.
• Degraded audio. On heavily accented, overlapping, or badly recorded speech, an experienced human transcriber can still beat the model by replaying, inferring, and asking clarifying questions.

For everyday speech-to-text, though, the conclusion is simple: on reasonably clean audio, Whisper matches or beats a paid human transcript at a fraction of the cost and delay.

Where Accuracy Drops: Noisy Audio, Meetings, Low-Resource Languages

Whisper's error rate climbs predictably as audio conditions get worse. The main culprits:

Distant and noisy microphones

Room echo, keyboard clatter, traffic, and air conditioning all raise WER. Benchmarks on poor-quality call-center audio show WER climbing to around 17.7% as of 2026, roughly six times the clean-audio figure for the same model family.

Multiple speakers and crosstalk

Whisper transcribes a single audio stream; it has no built-in speaker separation. When two people talk over each other in a meeting, the model has to guess, and it often drops or merges words. If transcribing recordings is your goal, see our guide to transcribing audio files on a Mac for realistic expectations.

Strong accents and specialized vocabulary

Whisper handles common accents well thanks to 680,000 hours of varied training audio, but rare accents, invented product names, and dense jargon still trip it up.

Hallucinations on silence and music

A known Whisper quirk: on long stretches of silence or background music, the model can occasionally invent text that was never spoken. This is another reason short, deliberate dictation is safer input than an hour-long ambient recording.

Why Dictation Is Whisper's Best-Case Scenario

Look back at the conditions that hurt accuracy: distant microphones, background noise, overlapping speakers, long unattended audio. Now look at what dictation involves:

• One speaker - you, with no crosstalk to untangle
• A close microphone - your Mac's built-in mic or a headset, centimeters from your mouth
• Deliberate speech - people dictate more clearly than they chat
• Short utterances - a sentence or a paragraph at a time, leaving no room for silence-induced hallucinations

Dictation is, in other words, close to what the benchmarks call clean audio, so a Whisper large model on a Mac operates near the top of the accuracy range. The "8-12% real-world WER" quoted online describes podcasts and phone calls, not a person speaking into their own laptop.

This is exactly the scenario Whisper Dictation for Mac is built around: hold Option+Space, speak a sentence, release, and the text appears at your cursor in any app.

Accuracy by Language: The Real Spread Across ~100 Languages

Whisper supports roughly 100 languages with automatic language detection, but "supports" and "excels at" are different things. Accuracy tracks the amount of training data per language, and the spread is large:

• Tier 1: English. The most training data and the best results, with clean-audio WER in the low single digits.
• Tier 2: major world languages. Spanish, French, German, Italian, Portuguese, Dutch, Japanese, Korean, and similar languages perform close to English-level accuracy on comparable audio.
• Tier 3: mid-resource languages. Usable results with noticeably more errors, especially on names and informal speech.
• Tier 4: low-resource languages. WER can exceed 25% as of 2026, meaning roughly one word in four is wrong. For these languages, Whisper is a rough draft tool at best.

Two practical implications: if you dictate in a major European or East Asian language, expect accuracy near English levels; and code-switching (mixing two languages mid-sentence) degrades results, so bilingual users get better output dictating each passage in a single language.

How to Get the Best Accuracy When Dictating on a Mac

Since audio conditions drive accuracy more than anything else, most of the gains are in your control:

• Use a large model. Large-v3 is the most accurate family released to date; smaller models are faster but make visibly more errors on accents and jargon. Dictation utterances are short, so a large model still feels quick on Apple Silicon.
• Keep the mic close. A built-in MacBook mic at normal working distance is fine; a headset mic is slightly better. Avoid dictating from across the room.
• Reduce steady background noise. Music with lyrics is the worst offender, since the model may transcribe the song instead of you.
• Speak in complete phrases. Whisper uses surrounding words as context, so a full sentence transcribes better than isolated words with long pauses.
• Let Whisper punctuate. It inserts punctuation on its own from your phrasing, one of its quiet strengths over older dictation systems.

You have two ways to run this on a Mac. The do-it-yourself route is installing whisper-cpp and wiring up your own hotkey, which we cover step by step in our guide to running Whisper locally on a Mac. The ready-made route is a native app that bundles the model, the hotkey, and the text insertion: our comparison of the best dictation software for Mac covers the main options, including Whisper Dictation.

Full disclosure on Whisper Dictation's own limits: it is macOS only (14.0 Sonoma or newer), it downloads a ~1.5GB model on first launch, and Intel Macs run it slower than Apple Silicon. In exchange, everything stays on your machine: audio is processed locally and deleted, with no account and no telemetry.

Frequently Asked Questions

The Bottom Line

Is Whisper accurate? Yes: state-of-the-art on the audio most people feed it, and better than human transcribers on clean speech. The honest picture is a range, from about 2.7% WER on clean benchmarks to 17.7% on poor call-center audio, and dictation sits at the good end because you control the microphone, the speaker count, and the pacing.

If you want to experience those numbers on your own voice instead of a benchmark chart, Whisper Dictation is a voice dictation app for Mac that runs Whisper 100% locally: one-time $9.99, works offline, 10 free transcriptions to test, and a 7-day money-back guarantee.