Do I need to buy a microphone for AI dictation?

Most people in Voibe's usage data did not. 331 of 507 dictated through the built-in microphone of their laptop or desktop, 15 used AirPods or EarPods, and each dedicated microphone model was used by fewer than ten people. Try the microphone you already own first.

Did you read people's dictations to produce the State of AI Dictation report?

No. Voibe has no transcript corpus to read. In on-device mode on Apple Silicon Macs, text never leaves the machine; in zero data retention cloud mode, audio is deleted the moment transcription completes. Every figure in the report is a count, a duration, a word count or an app category, aggregated across a small group of users who opted in to share usage data.

Is 507 people Voibe's whole user base?

No. The 507 people are the members of a small opt-in group who dictated at least once in the month the report covers. The report describes that group only, not Voibe's full user base and not dictation users in general.

Claude Takes 16% of Every Dictation and 24% of Every Word: The State of AI Dictation Report

Q: Which app do people dictate into most?

Claude. In Voibe's State of AI Dictation data, the Claude desktop app took 14,471 dictations and 557,153 words, 16% of all dictations and 24% of all words, from 122 people. That is more dictations than every email and chat app combined. All AI assistant apps together took 22% of dictations and 31% of words.

Q: Why do AI assistants get longer dictations than email?

Because detail gets cheap when you speak. A dictation into an AI assistant app averaged 38.6 words against 17.1 words for email, a 2.3× gap. When a paragraph takes about 20 seconds to say, people include the context, constraints and examples they would skip when typing.

Q: Is dictation really 3× faster than typing?

In Voibe's usage data, the median dictation ran at 122 words per minute and the median person at 107 wpm, measured on English dictations of at least 20 words and 5 seconds. Against the commonly cited 40 wpm typing average that is 3.05× and 2.7×; against a fast 80 wpm typist it is 1.5×. The figure includes pauses while the key is held.

Q: How much does Voibe cost?

Voibe costs $7.50 a month, $59 a year or $149 for a lifetime license, with a 7-day free trial and a 30-day money-back guarantee. Teams pricing is $6 per seat per month or $49 per seat per year.

The single biggest thing people dictate into is Claude.

Not email. Not Slack. Not a document. In one month, 14,471 dictations and 557,153 words went into the Claude desktop app. That's 16% of every dictation and 24% of every word. More dictations went into Claude than into every email and chat app combined.

Add ChatGPT and the other assistants, and AI apps took 22% of all dictations and 722,675 words, 31% of everything.

Typing was built for humans writing to humans. Voice is winning where the reader is an AI, because an AI wants context, and context is expensive to type.

The quick version:

16% of all dictations go into Claude. 24% of all words.
38.6 words per dictation to an AI. 17.1 to an email. People say twice as much to a machine.
22% of all dictations go into AI assistant apps. Add the editors and terminals where coding agents run and it's 46%.
The median dictation is 15 words and 8 seconds. Nobody dictates a document. They dictate a sentence, a hundred times a day.
122 words per minute spoken. The average typist does 40.
Developers are 17% of people and 24% of dictations. Keyboard people became voice people.
Dictations got 28% longer between March and August.
331 of 507 people use the mic in their laptop. 91 of 269 ever said “comma”. The friction is gone.

One month, 507 people, 89,791 dictations, 2,324,885 words, 428 apps, all aggregate.

Number	What it says about how work is changing
Claude: 16% of dictations, 24% of words	One AI app is the biggest destination for spoken text
38.6 words to an AI vs 17.1 to email	People brief machines in paragraphs and people in sentences
22% of dictations to AI assistant apps	The prompt is now a first-class thing people dictate
17% of people, 24% of dictations	Developers dictate more than anyone; the agent is the bottleneck now, not the keyboard
122 wpm vs 40 wpm typing	Detail is cheap when you speak it
20.3 to 25.9 words per dictation since March	Trust the transcript and you say more per press
331 of 507 on a built-in mic; 91 of 269 ever said “comma”	The hardware and the commands stopped mattering

Claude Is the Biggest Thing People Dictate Into

One app took 16% of every dictation. It wasn't Gmail, Slack or Word. It was Claude.

14,471 dictations into the Claude desktop app. 16.1% of the month.
557,153 words. 24% of everything dictated.
122 people, nearly one in four, dictated into Claude at least once.
38.5 words per dictation into Claude. The overall average is 26.

For scale: email took 6% of dictations and chat apps 5%. Claude alone beat both combined.

ChatGPT's desktop app came second among assistants: 3,521 dictations, 130,805 words, 41 people. Every AI assistant app together: 19,567 dictations (22%), 722,675 words (31%), 161 people.

A year ago the pitch for dictation was emails and documents. The data says the pitch is now the prompt.

Claude is the single largest app. Only two whole categories, code editors and browsers, are bigger.

Key Takeaway

Claude took 16% of all dictations and 24% of all words, more dictations than every email and chat app combined.

People Say Twice as Much to an AI as to a Person

Animation of two text fields filling with word-shaped blocks side by side. The left field, labelled Into an AI assistant app, fills to 39 blocks as a counter climbs to 38.6 words per dictation. The right field, labelled Into an email, fills to 17 blocks as its counter stops at 17.1. Three smaller figures then land: code editor or terminal 29.7 words, browser 24.3, chat app 18.3. The closing line reads: The assistant gets the paragraph. The colleague gets the sentence. Talk to machines, type to humans. — Average words per dictation, by where the text landed.

Where the text lands changes how much people say:

Into an AI app: 38.6 words per dictation
Into a code editor or terminal: 29.7
Into a browser: 24.3
Into a chat app: 18.3
Into an email: 17.1, and the quickest at 11 seconds median

That's a 2.3× gap between a prompt and an email. The assistant gets the paragraph. The colleague gets the sentence.

It makes sense once you think about who's reading. A colleague fills in the gaps. A model doesn't. So you give it the background, the constraints, the example. Typed, that's four minutes. Spoken, it's one. How to voice-prompt ChatGPT, Claude and Cursor covers the technique.

AI Assistant Apps Take 22% of All Dictations

AI assistant apps took 22% of all dictations and 31% of all words. Only two whole categories are bigger: code editors and terminals at 24%, and browsers at 23%.

The code editor number is worth a look too. Not every editor dictation is a prompt, but that's where Claude Code, Cursor and Warp live. Count them with the AI apps and you get 46% of dictations going somewhere software reads them first.

Where the text landed	Share of dictations	Average words per dictation
Code editors and terminals	24%	29.7
Browsers	23%	24.3
AI assistant apps	22%	38.6
Writing and notes apps	8%	—
Email	6%	17.1
Chat and messaging	5%	18.3
Everything else, including Voibe's own onboarding window (3%)	12%	—

Browsers reach the most people: 288 of 507 dictated into one. Browser dictations are shorter, and we don't look inside the tab.
The long tail is 428 different apps: spreadsheets, slides, design tools, video editors, project trackers.
The median person dictates into 3 apps. 137 people (27%) use five or more.

Dictation is a system-wide habit, not a feature inside one product. One key that works in every text field beats a microphone button inside one app. That's the case we make in dictating in Claude Code and dictating in ChatGPT.

Three categories take 69% of all dictations: code editors, browsers and AI assistants.

Nobody Dictates a Document. They Dictate an 8-Second Sentence.

Animation of three dictation bursts in a row. In each, a key pill reads HOLD while a waveform grows for 8 seconds, then flips to RELEASE as 15 word-shaped blocks land in a text field, followed by the note that the next press comes within 60 seconds for half of all dictations. A side panel stacks the figures: median dictation 15 words and 8 seconds; 50% followed by another within 60 seconds; 78% within 5 minutes; the 3.4% of dictations over 120 words carry 25% of words. The closing line reads: The 8-second sentence. Dictation is a burst, not a monologue. — Hold, speak a sentence, release, repeat. The word blocks are placeholders, not transcript text.

The median dictation is 15 words and 8 seconds. The average is about 26 words and 16 seconds. Even the 90th percentile is only 61 words and 32 seconds.

Nobody is dictating a memo in one go. People type with their voice, one sentence at a time.

The gaps between dictations show it:

Half of all dictations are followed by another within 60 seconds.
78% are followed by another within five minutes.

Hold the key, say a sentence, let go, read it, hold again. That is the 8-second sentence. It changes what a dictation tool should be built for: a key you can press a hundred times a day, text that lands before you look away, and punctuation you never have to say.

3.4% of Dictations Carry 25% of the Words

Short dictations are most of the presses. Long ones are most of the words.

Dictation length	Share of dictations	Share of all words
1–5 words	17%	2%
6–15 words	34%	12%
16–30 words	24%	19%
31–60 words	15%	22%
61–120 words	7%	20%
121–300 words	3%	19%
Over 300 words	0.4%	6%

The 3.4% of dictations over 120 words carried 25% of everything dictated. Those are the paragraphs, and most of them went to an AI. Voibe caps a single press at five minutes. 79 dictations hit that cap.

Half of all dictations are 15 words or fewer. They add up to 14% of the words.

Why Voice Is Winning the Prompt: 122 Words per Minute vs 40

The median dictation ran at 122 words per minute. The middle half ran between 95 and 151.

Per person, the median speaker did 107 wpm, with the middle half between 94 and 125. That's across 153 people with at least 20 qualifying dictations.

For comparison:

Wikipedia puts the average computer user at about 32.5 wpm and professional typists at 43 to 80 wpm. 40 wpm is the number most people quote.
Against 40 wpm, the median dictation is 3.05× faster and the median person 2.7×.
Against a fast 80 wpm typist, dictation is still 1.5× quicker.

122 wpm is slower than normal speech, which runs 150 to 160. The clock keeps running while you think. So this is the real-world number, pauses included.

That gap is the whole reason the prompt is going to voice. A 150-word brief to an agent costs four minutes typed and a little over one spoken. Speed is measured on English dictations of at least 20 words and 5 seconds, inside a 30 to 300 wpm window. 29,038 dictations qualified.

Median dictation speed, thinking pauses included, against the two typing benchmarks.

What That Adds Up to in a Month

2,324,885 words took 397 hours to say.

Typed at 40 wpm, the same words take about 969 hours.

That's 572 hours saved in one month, across 507 people. Roughly 24 days of nonstop typing. Want your own baseline? The typing speed test takes a minute.

The Developer Paradox: 17% of People, 24% of Dictations

Developers were supposed to be keyboard people. They dictate more than anyone.

85 people (17%) dictated into code editors and terminals.
They produced 21,717 dictations, 24% of the month.
About 255 dictations each. People dictating into AI apps did about 121 each. Twice the rate.
VS Code: 9,044 dictations from 38 people. Terminal: 1,765 from 21. Cursor and Warp both show up.

The keyboard was never the bottleneck for a developer typing code. It is the bottleneck for a developer explaining what they want to an agent. That's the job now, and it's a talking job.

Categories are by app, not by person. A developer who dictates a prompt into Claude counts under AI assistants for that one, and under code editors for the next. Many people sit in both columns. If that's your day: the developer dictation roundup and dictating in your terminal.

Dictations Got 28% Longer Since March

Average words per dictation, by month:

Month, 2026	Average words per dictation
March	20.3
April	22.8
May	24.2
June	23.9
July	26.1
August	25.9

That's 20.3 to 25.9, up 28% in five months. Words per active person per month went from about 4,225 to 4,586, up 8.5%. People are saying more per press, not just pressing more often.

Three things happened in that window, and all of them probably contribute:

We shipped cloud dictation. Transcripts got noticeably better, and people speak longer sentences to a tool they trust to get them right.
More prompts in the mix. AI assistant apps take the longest dictations. As their share grows, the average rises.
Habit. The longer you've used dictation, the less you stop to check.

Up 28% since March. Cloud dictation shipped in this window, and prompts to AI assistants grew as a share.

The Friction Is Gone

Two things used to stand between people and dictation: the microphone and the commands. Both have stopped mattering.

331 of 507 People Use the Mic That Came With the Laptop

65% of people dictate through the built-in microphone. MacBook Pro 153, MacBook Air 105, and the rest iMac, Studio Display and Windows laptop arrays.

AirPods or EarPods: 15 people
Blue Yeti: 6
DJI Mic Mini: 4
Razer, fifine, Jabra headsets: fewer than ten each

A Whisper-class model handles a modern laptop mic in a quiet room without help. Distance from your mouth matters more than the price of the capsule. Hundreds of people worked that out on their own. If you're about to spend $100 on a USB mic before trying dictation, don't. The best microphone for dictation explains when a better mic actually helps.

Dedicated microphones are a rounding error.

91 of 269 People Ever Said “Comma”

Animation of two rows of word-shaped blocks. In the top row, labelled the old way, red pills reading comma and full stop land between the blocks as the sentence is dictated. In the bottom row, labelled what people do now, the punctuation marks appear on their own in green. A panel then shows: 91 of 269 people ever spoke a punctuation command; 4% of their dictations contained one; 2,687 spoken commands in the whole month. The closing line reads: Nobody says comma anymore. Punctuation is automatic. — Spoken punctuation commands, by the numbers.

Of the 269 people on app builds that report it, 91 (34%) used a spoken punctuation command at least once. Those commands appeared in 4% of their dictations.

2,687 spoken commands in the whole month
1,316 punctuation marks
1,013 symbols
The rest: brackets, quotes and bullet commands

These are feature events counted by the app, not words read out of transcripts.

Automatic punctuation has won. Saying “comma” is now something people do for a symbol the model can't guess, not for the comma itself.

Smart Formatting: 208 On, 45 Off

Smart Formatting is Voibe's bounded clean-up pass. 208 people keep it on. 45 switched it off and dictated 8,325 raw transcripts in the month.

Most people want the clean-up. A firm minority wants exactly what they said and nothing else.

57% Zero-Retention Cloud, 43% On-Device

Of the 34,162 dictations tagged with an engine, 19,385 (57%) ran on Voibe's zero data retention cloud and 14,768 (43%) ran on the Mac itself, across eight quantised Whisper builds. Older app versions don't tag the engine, so this covers 38% of the month.

Cloud: 130 people.
Most-used local model: English-only Whisper large-v3-turbo, 5-bit. 6,875 dictations from 57 people.
Next: the English small model at 8-bit and 5-bit.
10 people run the multilingual large model locally.

Part of the cloud share is structural. Every Windows and Intel Mac dictation is cloud by design. On Apple Silicon it's a choice. Cloud vs local dictation covers the trade-off.

Where the transcription happened, for the 38% of dictations that report it.

Where to Start

Three places where the data says dictation pays off fastest:

You write prompts for Claude, ChatGPT or Cursor all day. Biggest single use, longest dictations: 38.6 words per press, 31% of all words. Start with the voice-prompting framework, or dictating in Claude Code if you live in a terminal.
You write code. The 85 people dictating into editors and terminals averaged 255 dictations each over the month. Dictating in VS Code is the quickest way to find out if it fits.
You've been waiting to buy a mic. Don't. 331 of 507 people use the one in their laptop, and a 2020 M1 is the second most common Mac among them. The getting started guide takes about ten minutes.

Voibe is the app all of this was dictated through. It runs on Mac and Windows. On Apple Silicon Macs it can transcribe fully on-device, so nothing leaves the machine. Everywhere else it uses Voibe's zero data retention cloud. $7.50 a month, $59 a year or $149 once, with a 7-day free trial and a 30-day money-back guarantee.

What to watch next: Claude's 16% share of dictations, AI apps' 22%, 25.9 words per dictation, 43% on-device. For a different kind of evidence, read one physician's year after leaving Dragon Medical One. For the technology behind the on-device share, why offline dictation matters.

Tip

Test dictation on the thing you do most, not on a paragraph you read aloud. For most people that's a 15-word message or a prompt to an AI. That's the test that predicts whether you'll keep using it.

Frequently Asked Questions About the State of AI Dictation Report

About the findings

Which app do people dictate into most? Claude. The Claude desktop app took 14,471 dictations and 557,153 words, 16% of all dictations and 24% of all words, from 122 people. That's more dictations than every email and chat app combined.

Why do AI assistants get longer dictations than email? Because detail gets cheap. A dictation into an AI app averaged 38.6 words against 17.1 for email. When a paragraph takes 20 seconds to say, people include the context and examples they'd skip when typing.

Is dictation really 3× faster than typing? The median dictation ran at 122 words per minute and the median person at 107 wpm. Against the commonly cited 40 wpm typing average, that's 3.05× and 2.7×. Against a fast 80 wpm typist, 1.5×. Measured on English dictations of 20+ words, pauses included.

Do I need a microphone for dictation? Most people didn't buy one. 331 of 507 dictated through the built-in laptop or desktop mic, 15 through AirPods or EarPods. Every dedicated microphone was used by fewer than ten people.

Did you read people's dictations to get this? No. There is no transcript corpus to read. Every figure is a count, a duration, a word count or an app category, aggregated across a small group of users who opted in.

About Voibe

Does Voibe work on Windows? Yes. Voibe runs on Mac and Windows. The Windows app (launched July 2026) uses Voibe's zero data retention cloud; the fully on-device mode is Mac-only (Apple Silicon).

How much does Voibe cost? $7.50 a month, $59 a year or $149 for a lifetime license, with a 7-day free trial and a 30-day money-back guarantee. Teams: $6 per seat per month or $49 per seat per year.

The Fine Print: Where This Data Comes From

A small subset of Voibe's users opted in, inside the app, to share usage data with us. This report is based on the data we collected from that group through the month of August 2026: 507 people who dictated at least once between 1 and 31 August 2026, 89,791 dictations, 2,324,885 words. The monthly trend section reaches back to March 2026 in the same data. It is not our full user base, and it's not a random sample of dictation users anywhere.

The data is a set of counters: timestamps, durations, word counts, the type of app the text landed in, device and input-device names, engine and model, and feature flags like Smart Formatting or a spoken-punctuation command. It reaches us only as totals across the whole group. There is no personal-level data in it, no per-person view for anyone at Voibe to open, and the app is built so it can't capture one.

What Voibe never collects, transmits or analyzes: in on-device mode on Apple Silicon Macs, audio and text never leave the machine. In zero data retention cloud mode, audio is transcribed on Voibe's own infrastructure by open-source models and deleted the moment transcription completes. Never stored, never sold, never used to train a model, and no third-party AI lab in the path (Voibe cloud privacy). No transcript content was read for this report. No window titles or browser tab contents were used. There are no per-user figures on this page, in the charts or in the animations. Japanese, Chinese and Korean dictations are left out of the words-per-dictation figures because the word counter undercounts them.