For dictation, anything under 10% WER on your own voice is workable, and under 5% feels excellent. For reference, professional human transcribers score around 5–6% WER on conversational speech, so a modern speech model in good conditions is already in human territory. The number that matters is the one you measure on your own audio, not a vendor's benchmark.
Learn · Updated July 2026
Voice typing accuracy explained: what word error rate really means, human vs AI error rates, and how to test and improve your speech-to-text
Every dictation tool advertises "industry-leading accuracy," and some claim 99%. This guide explains how voice typing accuracy is actually measured, what real benchmarks show for models like Whisper (and for professional human transcribers), why marketing percentages rarely match your experience — and the seven levers you control that matter more than which app you pick.
Start 14-day free trialHow accuracy is measured
Word Error Rate (WER): the metric behind every accuracy claim
The speech-recognition industry measures accuracy with Word Error Rate (WER): take a perfect reference transcript, compare the machine's output to it, and count three kinds of mistakes:
- Substitutions — the wrong word ("their" for "there," "cube cuddle" for "kubectl")
- Deletions — a word the system dropped entirely
- Insertions — a word the system added that you never said
WER = (substitutions + deletions + insertions) ÷ total words in the reference. Lower is better: 5% WER means one word in twenty is wrong, and "95% accuracy" is just 100 minus WER. If you see other terms on this page you don't recognize, our voice typing glossary covers WER, VAD, on-device vs cloud, and more.
Two things WER deliberately ignores are worth knowing. First, punctuation and casing usually don't count — a transcript with no commas and no capital letters can score a perfect WER while being miserable to read. Second, all errors weigh the same — mishearing "not" (which flips your meaning) counts exactly as much as mishearing "um." That's why two tools with identical WER can feel very different in daily use, and why the sections below separate raw model accuracy from output quality.
Real numbers
What benchmarks actually show in 2026 — humans included
| Condition | Typical WER | Word accuracy |
|---|---|---|
| Top model, clean scripted audio (e.g. Whisper large-v3 on LibriSpeech clean) | ~2.7% | ~97% |
| Top model, crowd-read sentences (Common Voice English) | ~5–6% | ~94–95% |
| Professional human transcribers, conversational telephone speech | ~5.1–5.9% | ~94–95% |
| Top models, mixed real-world recordings | ~8–12% | ~88–92% |
| Noisy, low-quality audio (e.g. call-center recordings) | ~15%+ | under ~85% |
Three takeaways from those numbers. First, the spread within one model is bigger than the spread between models — the same engine can be 3% or 15% WER depending purely on the audio you feed it. Second, machines reached human territory years ago: Microsoft's research systems matched the measured 5.1% human error rate on conversational speech back in 2017, and today's models are stronger. Third, dictation is the favourable case — you speak deliberately, close to a microphone you chose, so a well-set-up dictation workflow should land near the top of that table, not the bottom.
About those "99% accurate" claims: 99% means 1% WER, which even the best systems only approach on clean benchmark audio. When a marketing page quotes a single accuracy figure, it's the best case, measured on audio that isn't yours. The honest answer to "how accurate is this tool?" is always "measure it on your own voice" — which takes ten minutes (see the test below).
The five factors
What actually determines your voice typing accuracy
1. Microphone and environment — the biggest lever
Built-in laptop microphones sit next to the fan, under your typing hands, and half a metre from your mouth — they capture keyboard clatter, fan whine, and room reverb along with your voice. Benchmark data shows the same model losing 5–10+ WER points as audio quality drops. A modest USB headset, a lapel mic, or earbuds with a good integrated mic routinely improves results more than switching software.
Environment is the other half: background conversations, music with lyrics, and echoey rooms all degrade recognition. You don't need a studio — just a mic near your mouth and a room without competing speech.
2. Accent fit — real, measured, and fixable
Speech models are only as fair as their training data, and audits have found real disparities: one widely cited study of five major ASR systems measured average WER of 35% for Black American speakers versus 19% for white speakers, and audits of accented English found gaps from about 2 up to nearly 12 WER points versus "General American." That's the model's failure, not the speaker's.
Practically: if one engine handles your accent badly, try another before giving up — accent fit varies more between engines than headline accuracy does. Custom dictionaries then mop up the specific words that keep getting mangled.
3. Vocabulary — where "accurate" tools fall apart
General conversation is what models are trained on; your niche isn't. Product names, medical and legal terms, code identifiers, and colleagues' names are where a "97% accurate" tool suddenly writes "cipher" for Cypher, "cube cuddle" for kubectl, and "on you" for Aanya. Jargon-heavy dictation can add several WER points on its own.
The fix is a custom dictionary: tell the system your recurring terms once and the problem disappears permanently. FluidVox goes a step further and auto-learns from your corrections — when you fix a transcript, it validates the correction with AI and applies it in future sessions, so accuracy compounds over weeks of use.
4 & 5. Speaking style, and language coverage
Speaking style: models handle natural, conversational pace best. Racing through sentences, trailing off at the end, or over-enunciating robotically all hurt. You don't need to perform — just speak as if explaining something to a colleague, and finish your sentences at full volume.
Language: English, Spanish, Mandarin and other high-resource languages get the lowest WER; less-resourced languages can run 15–25%. Check that your tool actually supports your language well rather than nominally — FluidVox supports 99 languages on-device and 46 on its cloud engine, but like every tool, accuracy varies by language.
Raw WER vs what you see
Why two tools with the same WER produce different-quality text
AI cleanup: the accuracy you perceive
Raw speech-model output is a faithful record of how people actually talk — which means filler words, false starts, and restarted sentences. Modern dictation tools run an LLM cleanup pass that removes "uh" and "um," repairs restarts, fixes punctuation and casing, and formats the result for its destination. Two tools with identical raw WER can produce very different text depending on this layer.
This is the practical gap between built-in dictation (Apple Dictation, Windows voice typing — capable recognition, no cleanup) and tools like FluidVox, which applies one of six styles — natural, casual, professional, code, notes, email — automatically per app. Our Apple Dictation comparison covers this in depth, and how AI dictation works explains the full pipeline.
On-device vs cloud: a small and shrinking gap
Cloud speech APIs held a clear accuracy lead for years because the biggest models needed server hardware. As of 2026 that gap has narrowed to a few WER points or less on major languages: Whisper large-v3 and NVIDIA's Parakeet models run comfortably on modern laptops and score close to cloud APIs on most benchmarks.
The trade is now mostly about privacy and connectivity rather than accuracy: on-device means your audio never leaves the machine and dictation works offline. FluidVox's Local plan is fully on-device (Whisper and Parakeet); the Pro plan adds a cloud engine if you want it. For most people, on-device is the right default.
Run your own test
How to test speech-to-text accuracy in 10 minutes
1
Pick a realistic passage
Choose or write 200–300 words that resemble what you actually dictate — including the names, jargon, and numbers you use daily. Don't test with a news article if you dictate code reviews.
2
Read it into each tool
Use your normal microphone, your normal room, your normal pace. Testing in artificially quiet conditions tells you nothing about Tuesday afternoon with the dishwasher running.
3
Count the errors
Compare output to the original and count every wrong, missing, or extra word. Errors ÷ total words = your WER. A document diff tool makes this fast; for 250 words, 12 errors ≈ 5% WER.
4
Retest after fixing the basics
Add your problem words to the tool's dictionary, switch to a headset mic, and run the passage again. Most people see a bigger improvement from these two changes than from changing apps.
How to improve your accuracy
Seven levers, in order of impact
- Use a real microphone. A modest USB headset or good earbuds beat every built-in laptop mic. This is the single highest-impact change for most people.
- Cut ambient noise. Competing speech is the worst offender — close the door on conversations and lyrics-heavy music. Steady fan noise matters less than voices.
- Speak at conversational pace, and finish your sentences. Trailing off at the end of a thought is a top cause of dropped words.
- Build your custom dictionary. Add the 30–50 terms that keep getting mangled — names, jargon, product names. With FluidVox this largely happens automatically via correction learning.
- Match the model to your accent. If results are poor, test a different engine before concluding voice typing "doesn't work for you" — accent fit varies between models.
- Use AI cleanup for the destination. Per-app styles (an email reads differently than a code comment) close the gap between "correct words" and "usable text."
- Pick the right tool for your platform. See our 2026 Mac comparison, Windows comparison, or start with what is voice typing? if you're new to this.
Voice typing accuracy FAQ
On clean English audio, top models like Whisper large-v3 report roughly 2–6% WER on standard benchmarks (94–98% word accuracy), but real-world recordings typically land in the 8–12% WER range. Your personal result depends heavily on microphone, environment, accent, and vocabulary — which is why the same tool can feel flawless for one person and frustrating for another.
On clear conversational English, roughly yes: research measured professional human transcribers at about 5.1–5.9% WER, and top AI systems reached the same level as far back as 2017. Humans still win on messy audio — heavy crosstalk, strong accents, poor recordings — and on judgment calls like formatting. For dictation, where you speak deliberately into a nearby mic, modern AI is usually accurate enough that the remaining difference comes from cleanup and vocabulary handling.
Pick a 200–300 word passage that resembles what you actually dictate, read it aloud into each tool with your normal mic and room, then compare the output to the original and count the errors. Errors divided by total words is your WER. Testing with your own voice, vocabulary, and equipment beats any published benchmark — vendor numbers are measured on audio that isn't yours.
Nine times out of ten it's the input, not the model: a laptop mic picking up fan and keyboard noise, background chatter, speaking too fast or trailing off at sentence ends, or specialist vocabulary the model has never seen. Fix the microphone first, then add your recurring problem words to a custom dictionary. If accuracy is still poor, try a different engine — accent fit varies between models.
Treat them skeptically. A 99% claim means 1% WER, which top models only approach on clean, scripted, benchmark-style audio — real-world recordings typically measure 8–12% WER even for the best systems. Marketing pages quote the best-case benchmark; your mileage is set by your mic, room, accent, and vocabulary. Run your own 200-word test instead of trusting the banner number.
There is no single winner — rankings shift with the test set, language, and audio conditions, and independent benchmarks regularly disagree with vendor claims. As of 2026, Whisper large-v3, NVIDIA's Parakeet family, and the top cloud APIs (Deepgram, AssemblyAI, Google, Azure) all cluster within a few WER points of each other on English. FluidVox runs Whisper and Parakeet on-device and adds a cloud engine on Pro, so you can test which works best for your voice.
Yes — it is usually the single biggest lever you control. Speech models degrade sharply as input audio gets noisier or more reverberant, and built-in laptop mics capture fan noise, keyboard clatter, and room echo. A modest USB headset or earbuds with a decent integrated mic routinely improves results more than switching software.
Yes, and it's well documented: audits of major speech systems found error rates several points higher — in some studies nearly double — for accents under-represented in training data. If your accent fares badly with one engine, try another before giving up; accent fit varies a lot between models. Custom dictionaries also compensate for the specific words that get mangled repeatedly.
The microphone matters more than the noise cancellation. Earbuds like AirPods help mainly because the mic sits close to your mouth, and their voice-isolation processing reduces ambient pickup before the audio reaches the speech model. A directional headset mic in a quiet room remains the gold standard.
The base speech model doesn't retrain on you, but tooling around it can learn. FluidVox auto-learns from your corrections — when you fix a mistranscription, it validates the fix with AI and applies it in future sessions — so perceived accuracy compounds over weeks. With tools that don't auto-learn, you have to maintain the dictionary by hand.
The gap has narrowed to a few WER points or less on major languages. Large on-device models like Whisper large-v3 run well on modern laptop hardware and score close to cloud APIs on most benchmarks, while smaller on-device models trade some accuracy for speed. FluidVox offers both: on-device Whisper and Parakeet on the Local plan, plus a cloud engine on Pro if you want it.
Free tools like Google Docs voice typing use capable engines and can be surprisingly accurate on clear speech — accuracy isn't the main thing you give up. What free web tools lack is everything around the transcript: custom vocabulary, auto-punctuation quality, filler-word removal, per-app formatting, offline privacy, and working outside one browser tab. If you dictate occasionally, free is fine; if you dictate daily, the workflow gaps cost more than the software.
No. Even professional human transcribers make roughly 5% errors on conversational speech, so 100% is not on the table for anyone. The right question is which tool gets closest for your specific voice, environment, and vocabulary — and how quickly it lets you fix and learn from the errors that remain.
Try FluidVox free for 14 days
Full access, no credit card required. Then $2.99/month or $39 one-time.
Start free trial