Voice dictation statistics (October 2026): speed, accuracy, models and limits, with sources
By Ryleigh Newman ยท Published
The most cited voice dictation statistic is a speed one: in a study by researchers at Stanford, the University of Washington and Baidu, people entered short English messages on an iPhone at 153 words per minute by voice and 52 on the keyboard, 2.93 times faster. Below are the other numbers I trust on speed, errors, the speech models, Apple's built-in limits and free plan caps, each with what it measures, who measured it, when, and a source you can check.
I built YapToText, a free Mac dictation app that runs on device, because my hands make typing difficult, and I dictate every day. That makes me an interested party, so this page follows one rule: no number appears unless its source states it, and every source is listed at the end with the date it was checked. I publish no accuracy percentage for YapToText or any model, so you will not find one here. For a number about you, the section on measuring your own shows how to get it on your Mac.
Download for free. No account, nothing locked, macOS 14 or later on Apple silicon. Source on GitHub.
On this page: 11 sections
Steps
- Dictate for a few days with YapToText: click where the words go, tap Right Command, talk, and tap it again.
- Open Statistics in the sidebar, or the chart button on the History page.
- Read your speaking pace in words per minute; a dash means you have spoken for less than 30 seconds in total.
- Read the time saved in Insights, which compares against typing at 40 words a minute.
- Divide words spoken by your own typing speed and subtract time speaking to redo time saved for yourself.
Speed statistics: speaking versus typing
Three university studies give the clearest speed numbers. Speech beat the on-screen keyboard by close to three times on a phone; none of the studies cited here measured a Mac.
Checked October 5, 2026, on arXiv, the ACM Digital Library and the University of Cambridge's own news page.
| Statistic | What it measures | Who measured it | Year |
|---|---|---|---|
| 153 vs 52 words per minute | Copying short English messages on an iPhone 6 Plus: speech with Baidu's Deep Speech 2 against the iOS keyboard | Ruan, Wobbrock, Liou, Ng and Landay (Stanford, University of Washington, Baidu) | 2016 on arXiv; journal version January 2018 |
| 2.93 times faster | The same English test, speech speed over keyboard speed | The same study | 2016, 2018 |
| 123 vs 43 words per minute, 2.87 times faster | Copying short messages in Mandarin Chinese, speech against the Pinyin keyboard | The same study | 2016, 2018 |
| 28 participants | Composing their own text and copying set text on a smartphone: speech was faster in both, by less when composing | Foley, Casiez and Vogel, CHI 2020 | 2020 |
| 52 words per minute | Average typing speed on a computer keyboard, on randomized phrases | Dhakal, Feit, Kristensson and Oulasvirta, reported by the University of Cambridge | 2018 |
| 136 million keystrokes from 168,000 volunteers | The size of that typing study | The same study | 2018 |
| 120 words per minute | The fastest participants in that typing study | The same study | 2018 |
| 60 to 90 words per minute | What professionally trained typists of the 1970s and 1980s typically reached | As stated in the University of Cambridge's story | 2018 |
Three details keep these numbers honest. The phone study had people copy messages they were given, so nobody had to decide what to say, and speech also made fewer errors during entry there. The composing study found that speech needed more preparation time before people started talking. And the 52 words per minute in the two studies is a coincidence of two different tests, one on a phone screen and one on a computer keyboard, not a comparison. The guide to dictation versus typing speed walks through each study and where dictation is still slower.
Accuracy and error statistics
Accuracy numbers are scarcer than speed numbers, and most come from the people who make the models. The independent number is about hallucinations; the word error rates are the makers' own.
Hallucinations: words nobody said
- Roughly 1%. Share of transcriptions that contained entire hallucinated phrases or sentences, with nothing like them in the audio. Measured by Koenecke, Choi, Mei, Schellmann and Sloane in "Careless Whisper: Speech-to-Text Hallucination Harms", presented at ACM FAccT 2024.
- 38%. Share of those hallucinations that included explicit harms, such as perpetuating violence, inaccurate associations or implied false authority. Same study, 2024.
- 13,140 segments, nearly 40 hours. The audio tested: speech from AphasiaBank, from speakers with aphasia and a control group, sent to OpenAI's Whisper API in 2023. Same study.
- 12 of 187. Audio segments that still produced hallucinations when the authors reran them in December 2023, after many earlier examples were resolved. Same study.
Those figures describe OpenAI's hosted API on that sample in 2023. They are not a rate for whisper.cpp, for Whisper Large v3 Turbo on a Mac or for YapToText. The Whisper hallucinations guide explains why they happen and what YapToText does about them.
Word error rate, as the makers report it
- 6.32% vs 7.44%. Average word error rate on English test sets from the Hugging Face Open ASR Leaderboard, such as LibriSpeech, AMI and Earnings-22, for NVIDIA's Parakeet TDT 0.6B v3 against Whisper large-v3. Lower is better: it is the share of words the model got wrong. Measured by NVIDIA, in its own paper, 2025.
- 50% fewer errors. OpenAI says that, measured zero-shot across many diverse datasets, Whisper makes 50% fewer errors than models specialized for the LibriSpeech benchmark, though it does not beat them on LibriSpeech itself. Stated by OpenAI when it introduced Whisper in 2022.
Both are a maker's measurement of its own model, on recordings that are not your voice. The Parakeet vs Whisper comparison lists the limits of NVIDIA's numbers, and the guide to improving dictation accuracy covers what actually changes the result for one person at one microphone. Apple publishes no accuracy rate for Dictation to put here, and I have not measured one for YapToText.
Statistics about the speech models
These are the size, training and language numbers behind the models most Mac dictation apps run. Model numbers describe the model; how it feels to dictate depends on the app around it.
Checked October 3 and October 5, 2026, on OpenAI's and NVIDIA's own pages and model cards, and in YapToText 1.5.3.
| Statistic | What it measures | Who stated it | Year |
|---|---|---|---|
| 680,000 hours | Whisper's training data: multilingual and multitask supervised data collected from the web | OpenAI | 2022 |
| About a third | Share of Whisper's audio dataset that is not English | OpenAI | 2022 |
| 809 million parameters | The size of Whisper Large v3 Turbo | OpenAI, on its model card | Checked 2026 |
| About 8 times | Turbo's speed relative to Whisper large, with minimal degradation in accuracy, measured on an A100 GPU, not a Mac | OpenAI, in the Whisper README | Checked 2026 |
| 600 million parameters | The size of Parakeet TDT 0.6B v3 | NVIDIA, on its model card | Checked 2026 |
| 25 European languages | The languages Parakeet TDT 0.6B v3 transcribes | NVIDIA, on its model card | Checked 2026 |
| 24 minutes, or 3 hours | Longest audio Parakeet v3 takes in one pass: 24 minutes with full attention on an A100 80GB GPU, up to 3 hours with local attention | NVIDIA, on its model card | Checked 2026 |
| RTFx 3,332.74 | Parakeet v3's throughput on the Open ASR Leaderboard: duration of audio transcribed divided by computation time | NVIDIA, in its paper | 2025 |
| 574 MB | Whisper Large v3 Turbo (Q5), the speech model that ships inside YapToText's App Store build | YapToText | 2026 |
| 29 languages, plus Auto-detect | The Language picker in YapToText 1.5.3, 32 entries in all | YapToText | 2026 |
OpenAI's model card lists Whisper's languages by name rather than as a headline count, and says performance varies widely depending on the language, so I give no single number for it. The Mac dictation languages page lists YapToText's 29, and YapToText's AI models guide lists every model the app offers with the size it shows: the App Store download is about 2.9 GB, most of it the two bundled models, the 574 MB speech model and Phi-3.5 Mini Instruct for cleanup, shown as 2.3 GB.
Apple's built-in limits, in numbers
Apple publishes a few hard numbers for the dictation built into every Mac. Apple Dictation has no length limit while you talk, and a 30-second limit when you stop.
Checked October 3 and October 5, 2026, in Apple's Mac User Guide.
- 30 seconds. Apple's Mac User Guide says Dictation stops automatically when no speech is detected for 30 seconds. In the same paragraph it says you can dictate text of any length without a timeout, so this is a silence cutoff, not a session limit.
- 1,000 terms. Voice Control, Apple's separate accessibility feature, lets you add up to 1,000 custom vocabulary terms and record how you say each one, in languages that support a custom vocabulary. Dictation itself has no word list.
- One at a time. Apple says that while Voice Control is on, standard Dictation is not available, so the 1,000-term vocabulary and Dictation are never in use together.
- Languages. Apple keeps its list of Dictation languages on its macOS Feature Availability page, and says Dictation is not available in all languages or regions and that its features vary. Apple's pages give the list, not a count, so this page does not give one.
The page on the 30-second limit explains what happens when you pause to think, and Voice Control vs Dictation covers when the 1,000-term vocabulary is worth the switch. For comparison, YapToText has no silence timeout by default: it listens until you tap the key again, with a safety stop that ends a single dictation at 30 minutes and still saves it.
What free plans allow
Free plans are where the most specific numbers live, and where they change fastest. Most cloud dictation apps cap free use by words or minutes; local apps mostly do not.
Checked October 3 and October 5, 2026, on each app's own pricing page or help center. Plans change, so check before you rely on one.
| App | Free allowance | Measured as | Checked |
|---|---|---|---|
| Wispr Flow | 2,000 words a week on a Mac or Windows PC; 1,000 on a phone | Words a week. Its help center says desktop dictation keeps working to 5,000 words, then stops until the weekly reset; phones stop at 1,500 | October 3, 2026 |
| Typeless | 8,000 words a week, at standard accuracy | Words a week | October 5, 2026 |
| Otter.ai Basic | 300 transcription minutes a month, up to 30 minutes per conversation, 3 file imports per account | Minutes a month | October 5, 2026 |
| Aqua Voice | 1,000 free words, which Aqua's Mac page puts at about 7 minutes of talking | Words, a trial | October 3 and 5, 2026 |
| Monologue | A trial of 1,000 dictation words and 10 recorded notes | Words, a trial | October 5, 2026 |
| Willow Voice Basic | Unlimited dictation on Frontier Mini, the model Willow calls weaker, 20 Scribe uses a week, 5 minutes per dictation | Uses a week and minutes per dictation; a help center chart also shows 2,000 words a week, which the pricing page does not | October 3 and 5, 2026 |
| superwhisper | Unlimited dictation with local Whisper models, up to two modes without AI processing | No word cap | October 5, 2026 |
| Spokenly | Local Whisper and Parakeet models with no word caps | No word cap | October 5, 2026 |
| VoiceInk | A 7-day trial, then a one-time license | Days | October 5, 2026 |
| Apple Dictation | Built in, no cap; stops after 30 seconds without speech | No word cap | October 3, 2026 |
| YapToText | Everything, with no word cap, no account and no paid tier | No word cap | October 3, 2026 |
Three things to read with the table. A weekly cap resets, a trial does not. A minute cap and a word cap are not the same unit, and I have not converted one into the other. And a free plan that runs in the cloud and one that runs on your Mac are not the same deal, whatever the number. The details and the paid plans are on the Wispr Flow pricing, Typeless pricing and Otter.ai pricing pages, and the best free dictation apps for Mac compares the free options side by side.
Measure your own dictation statistics
Every number above describes someone else. YapToText's Statistics page adds up your own dictation from your History, on your Mac. Your own words per minute is the statistic that settles whether dictation is faster for you. YapToText is free on the Mac App Store and needs an Apple silicon Mac with macOS 14 or later.
Step 1: Dictate for a few days
- You: Click where the words go, tap Right Command, talk, and tap it again. Do this through ordinary email, messages and notes.
- YapToText: Files each dictation in History, with the words and how long you spoke.
- Check: Your dictations appear on the History page.
Step 2: Open Statistics
- You: Click Statistics in the sidebar, or the chart button on the History page.
- YapToText: Shows four totals, dictations, words spoken, time speaking and your speaking pace in words per minute, then a bar chart of words per day for the last 30 days.
- Check: Speaking pace shows a number. A dash means you have spoken for less than 30 seconds in total, so keep dictating.
Step 3: Read the insights
- You: Look at the Insights card.
- YapToText: Shows the time saved versus typing everything at 40 words a minute, your current streak and your best one, and how many words a typical dictation runs.
- Check: Remember that time saved counts talking time only, not the time you spend reading and fixing.
Step 4: Redo time saved at your own typing speed
- You: Divide words spoken by your typing speed from any typing test, then subtract time speaking.
- Check: As arithmetic only: 3,000 words at 40 words a minute is 75 minutes of typing, and at 60 words a minute it is 50.
Statistics follow History, so the defaults, Keep last 500 and auto-delete after 30 days, trim the totals, and with history off, new dictations are not counted. Every number is worked out on your Mac when you open the page, and nothing is tracked or sent anywhere. The Statistics guide describes every card.
How to cite this page
You are welcome to cite this page, but the stronger citation is almost always the original source. Cite the study for the number, and this page for finding it.
- Quote the source's own words. Each number here keeps the source's wording and precision: roughly 1% stays roughly 1%, and 2.93 times is not rounded to three.
- Keep the context. Say what was measured and on what: an iPhone in a lab, OpenAI's API in 2023, NVIDIA's English test sets.
- Give the date. Free plans and limits were checked on the dates shown and may have changed since.
Cite as: a plain reference for this page.
Newman, R. (2026, October 5). Voice dictation statistics (October 2026). YapToText. https://yaptotext.com/dictation-statisticsWhat happens: A reader can find this page, and from its Sources list, the original study behind each number.
If a number here no longer matches its source, open an issue on GitHub. Every one gets read.
What these numbers cannot tell you
- Your speed on a Mac. The best-known speed figure comes from an iPhone, set phrases and a 2016 recognizer. No study I found compared dictation with a full keyboard on a desktop computer.
- Your accuracy. Word error rates were measured by the makers on benchmark recordings. Your voice, microphone and vocabulary are not in those test sets.
- Today's hallucination rate. The Careless Whisper figures describe one hosted API on one sample in 2023, not every Whisper setup now.
- Anything about YapToText's accuracy. I publish no accuracy percentage for any model.
- Tomorrow's free plans. Caps and prices are a snapshot from the dates shown.
Published numbers show what is possible; only your own numbers show what happens for you.
The short version
- Speed. 153 against 52 words per minute, 2.93 times faster, by voice on an iPhone, Ruan and colleagues, 2016 and 2018. The average keyboard typist does 52 words per minute, Cambridge and Aalto, 2018.
- Errors. Roughly 1% of transcriptions with hallucinated phrases, 38% of those harmful, Careless Whisper, 2024. Word error rates of 6.32% and 7.44% are NVIDIA's own figures.
- Limits. Apple Dictation stops after 30 seconds without speech; Voice Control holds up to 1,000 terms. Wispr Flow's free plan is 2,000 words a week, Typeless's 8,000, Otter's 300 minutes a month.
- Your own. YapToText's Statistics page shows your speaking pace and time saved, worked out on your Mac.
Cite the original source, keep its context, and measure yourself before you believe any multiplier.
If you know of a study this page should include, or a number here has changed at its source, open an issue on GitHub. Every one gets read.
Download for free. No account, nothing locked, macOS 14 or later on Apple silicon. Source on GitHub.
Related
Sources
- https://arxiv.org/abs/1608.07323 (checked 2026-10-05)
- https://dl.acm.org/doi/10.1145/3161187 (checked 2026-10-05)
- https://dl.acm.org/doi/10.1145/3313831.3376861 (checked 2026-10-05)
- https://www.cam.ac.uk/research/news/what-makes-a-faster-typist (checked 2026-10-05)
- https://dl.acm.org/doi/10.1145/3173574.3174220 (checked 2026-10-05)
- https://arxiv.org/abs/2402.08021 (checked 2026-10-05)
- https://dl.acm.org/doi/10.1145/3630106.3658996 (checked 2026-10-05)
- https://news.cornell.edu/stories/2024/06/ai-speech-text-can-hallucinate-violent-language (checked 2026-10-05)
- https://arxiv.org/abs/2509.14128 (checked 2026-10-05)
- https://blogs.nvidia.com/blog/speech-ai-dataset-models/ (checked 2026-10-05)
- https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 (checked 2026-10-05)
- https://openai.com/index/whisper/ (checked 2026-10-05)
- https://huggingface.co/openai/whisper-large-v3-turbo (checked 2026-10-05)
- https://github.com/openai/whisper (checked 2026-10-03)
- https://github.com/openai/whisper/blob/main/model-card.md (checked 2026-10-05)
- https://support.apple.com/guide/mac-help/use-dictation-mh40584/mac (checked 2026-10-03)
- https://support.apple.com/guide/mac-help/use-a-custom-vocabulary-mchl3eb7b79a/mac (checked 2026-10-05)
- https://support.apple.com/guide/mac-help/turn-voice-control-on-or-off-mchl63d14732/mac (checked 2026-10-03)
- https://www.apple.com/macos/feature-availability/ (checked 2026-10-03)
- https://wisprflow.ai/pricing (checked 2026-10-03)
- https://docs.wisprflow.ai/articles/3715544288-weekly-word-limit-not-resetting-troubleshooting-guide (checked 2026-10-03)
- https://www.typeless.com/pricing (checked 2026-10-05)
- https://help.otter.ai/hc/en-us/articles/360047538094-Conversation-import-and-app-limits-on-the-Basic-free-plan (checked 2026-10-05)
- https://aquavoice.com/pricing (checked 2026-10-05)
- https://aquavoice.com/mac (checked 2026-10-03)
- https://www.monologue.to/pricing (checked 2026-10-05)
- https://willowvoice.com/pricing (checked 2026-10-05)
- https://help.willowvoice.com/en/articles/12854184-willow-pricing-plans-overview (checked 2026-10-05)
- https://help.willowvoice.com/en/articles/12772803-how-to-change-your-dictation-time-limit-in-willow (checked 2026-10-03)
- https://superwhisper.com/docs/billing/plans (checked 2026-10-05)
- https://spokenly.app/pricing (checked 2026-10-05)
- https://tryvoiceink.com/pricing (checked 2026-10-05)
- https://github.com/ryleighnewman/YapToText (checked 2026-10-03)
- https://apps.apple.com/us/app/yaptotext/id6786382289 (checked 2026-10-03)
Questions
How much faster is dictation than typing?
In a study by Stanford, the University of Washington and Baidu, people copying short English messages on an iPhone were 2.93 times faster by voice, 153 words per minute against 52 on the keyboard. That was on a phone, and none of the studies cited here measured a Mac.
What is the average typing speed?
About 52 words per minute, from a 2018 study of 136 million keystrokes by 168,000 volunteers reported by the University of Cambridge. The fastest participants typed 120 words per minute.
How accurate is speech to text?
It depends on the audio and the model. NVIDIA's own paper reports an average word error rate of 6.32% for Parakeet TDT 0.6B v3 and 7.44% for Whisper large-v3 on English test sets, but those are a maker's numbers on benchmark recordings, not your voice.
How often does Whisper hallucinate?
The Careless Whisper study, presented at FAccT 2024, found roughly 1% of transcriptions from OpenAI's Whisper API in its 2023 sample contained invented phrases or sentences, and 38% of those included explicit harms.
How long can you dictate on a Mac?
Apple's guide says you can dictate text of any length while you keep talking, but Dictation stops automatically after 30 seconds without speech. YapToText has no silence timeout by default, with a safety stop at 30 minutes per dictation.
How can I measure my own dictation speed?
In YapToText, open Statistics in the sidebar. It shows your speaking pace in words per minute and the time saved against typing at 40 words a minute, worked out on your Mac from your History.