The quiet comeback of voice

The assistants stalled. Dictation took off — by its own account

Say “Alexa, set a timer” and you are using the first kind of voice input. Dictate an email instead of typing it, and you are using the second.

There are five kinds in all, and over the last two years they went different ways: commands to assistants, dictation of text, voice messages, spoken conversation with AI models, and voice for programming and for giving tasks to AI agents.

The first kind has stopped growing in reach. In the United States, 33% of people owned a smart speaker in 2021 and 35% in 2025; the survey firm itself says the number barely moved in four years.¹

Britain is the exception: from 25% to 45% over the same years, mostly thanks to Alexa.² Use per owner kept rising for longer: tasks per week on a speaker went from 7.5 in 2017 to 12.4 in 2022.³

Commercially, though, the first wave failed. Amazon's devices unit lost more than $25 billion between 2017 and 2021, because people used Echo for alarms and weather, not for shopping.⁴

The second wave is visible almost only in what companies say about themselves. Wispr Flow, a dictation app, went from a $700 million valuation in November 2025 to $2 billion in August 2026.⁵

Google says 63% of the Gemini app's more than one billion monthly users “use voice”, without saying how that was counted.⁶ Apple says Siri requests grew from about 1.5 billion a day in 2024 to more than 2.5 billion in 2026.⁷

The big independent surveys — Pew, YouGov, Edison, Ofcom — do not yet ask about dictation or voice chat with AI as a separate question. The closest one found that 57% of Americans aged 12 and over had used a generative AI assistant at least once, typed or spoken.⁸

So the growth curve of the second wave is drawn by interested parties.

Voice messages, the third kind, are a mass habit but not a growing one. WhatsApp's last official count was 7 billion a day, in March 2022.¹⁰

In a YouGov survey of 2,442 Americans in February 2026, about 10% used voice messages several times a week and 41% said the format did not apply to them. As many used them less (12%) as more (11%).⁹

Three times faster at copying. About 1.4 times at composing

Here is the most quoted number about voice: speaking is three times faster than typing. It is true — for copying. In the experiment behind it, 48 people, 24 per language, copied short ready-made phrases on an iPhone: 153 words per minute by voice against 52 on the on-screen keyboard in English, 123 against 43 in Mandarin.¹¹

The authors themselves call it an upper bound reached in a lab. And nobody writes by copying.

In 2020 Foley, Casiez and Vogel asked 28 people to compose a sentence of their own on a phone, linking three simple pictures. While words were actually being spoken or typed, voice still ran three times faster: 116.5 against 35.1 words per minute.

But counted over the whole attempt — thinking, speaking, fixing — a message took 20.8 seconds by voice and 29.2 by keyboard. That is about 1.4 times, not three.¹²

Two things ate the difference. Before the first word, people thought twice as long when they were about to speak: 11.0 seconds against 5.3. Typing, people think as they go; speaking, they plan the sentence first.

And any correction was expensive: across both methods and both tasks, attempts that needed a fix ran at 54.6 words per minute, against 141.1 for clean ones.

For scale, volunteers on an online typing test averaged 52 words per minute on a physical keyboard, across 168,000 people,¹³ and 36 on a phone, across 37,000.¹⁴

There is no controlled comparison of composing by voice against a desktop keyboard with today's speech recognition. The last one, from 1999, had first-time users composing by voice at 7.8 words per minute, corrections included, against 19 by keyboard and mouse; copying text, it was 13.6 against 32.5.¹⁵

Your speaking speed is how fast your mouth goes, not how fast dictation goes: nothing is recognised here, and nothing needs correcting. The ratio between your two numbers is the most voice can ever give you.

In Foley's experiment, speaking alone was 3.3 times faster than typing, and the whole attempt 1.4 times: after thinking and fixing, about 40% of the speed ratio survived.

What the machine hears depends on who is speaking

Say “move the meeting to Thursday at ten” in a quiet room with a newsreader's accent, and a good recogniser will get it right. Say it with a Lagos accent over café noise, and it may type “thirsty at then”.

Recognition is scored by word error rate: the share of words it gets wrong, drops or adds. In October 2025 the best model on a standard public leaderboard averaged 5.63% across its test sets.

Whisper large-v3, the open model behind many dictation apps, averaged 7.44%.¹⁶ For comparison, careful professional transcription was measured at 5.1% on recorded phone conversations in 2017.¹⁷

Those averages hide who is speaking. On a collection of conversational English with accents from around the world, Whisper got 19.7% of words wrong, against 2.7% on clean read American English. For Nigerian and Jamaican English even the stronger models went above 20%, and an academic model above 45%.¹⁸

In 2020, five commercial systems got 35% of words wrong from Black speakers and 19% from white speakers: the same system, almost twice the errors.¹⁹

Generative recognition added a new kind of failure: making things up. Whisper invented whole phrases in about 1% of transcripts, and 38% of those inventions contained something harmful, such as violence or false claims.²⁰

Long pauses, common in people with aphasia, made inventions more likely. An invented sentence is worse than a misheard word, because it reads smoothly.

The bottleneck has moved from recognition to correction. Editing by voice has two unsolved problems: telling a command from text that should be written down, and pointing at the place to fix.

Tapping near the error while saying the correction, and letting a language model work out what goes where, cut task time by 16.4% and keyboard taps by 47.5% against the previous best method, in a test with 16 people.²¹

Dictated text also comes out wordy and loose, which is why the new apps rewrite it with a language model before pasting. In a diary study of 7 to 10 days with such a tool, 9 of 12 writers adapted to it, with a median comfort of 5 out of 7.

Those who kept their doubts named the old hard parts: thinking of the next sentence while speaking, and putting the cursor where it should be.²²

Voice speeds up the mouth, not the planning

If voice is three times faster, why doesn't writing get three times faster? Because most of writing is not typing. A 1983 study of letter writing found that about two thirds of composing time is planning — between 58% and 78% across the setups tested.

Letters came out equally good whether people wrote them by hand, dictated or spoke them.²³ A faster output channel speeds up only the remaining third.

Ben Shneiderman proposed the mechanism in 2000: speaking and short-term memory draw on the same resource, while moving a hand runs in parallel with thinking.

In one of his lab's experiments, spoken commands were 12–30% faster than the mouse, but saying them aloud made it harder to remember symbols that had to be typed back in later.²⁴

Later studies line up along one line, from copying to solving. Where the job is copying, voice wins on speed and feels easier. Where the job is composing, voice is a little faster and feels the same.

Where the job is solving a problem, voice loads the mind more. For 50 people using a medical chat assistant, speech raised measured mental workload on both hard and simple tasks.²⁵ 35 emergency doctors documenting by voice were 18.1% slower and made 138 errors against 32.²⁶

The clear winners are people whose bottleneck is the hand or the spelling. Students with learning disabilities wrote better essays with speech recognition than by hand; for other students it made no difference.²⁷

Talking to AI sits on the same line. In the largest study I found, 919 students in an introductory programming course could choose how to prompt an AI tool. Almost nine in ten chose text only (88.5%), 4.8% voice only and 6.7% both.

Unedited voice prompts worked on the first try less often: 35% against 52% on the first task and 38% against 58% on the second. Once students edited the transcript, the gap was no longer significant.²⁹

In creative tasks, voice gave a faster first prompt and more, longer follow-ups, but changed neither the effort people felt nor how they rated the results.²⁸

The popular idea that thinking out loud helps holds only halfway. Across 94 studies, simply voicing your thoughts did not change accuracy and only made the work slower; accuracy rose when people had to explain their reasoning.³⁰

Explaining a problem to an AI model by voice may well help. Nobody has tested it.

The voice tires by the dose. Nobody has measured the dose of dictation

Typing all day tires your hands. Talking all day tires something else, and there is no scientific daily limit for how much of it is safe. What is measured well is how much voice professionals use.

Teachers, a job with high rates of voice disorders, sound their vocal folds for 23% of working time, against 13% off work and 12% at weekends, in a study of 31 teachers from 2007.³¹ A newer group of 26 teachers came out at 16.2%, 8.4% and 8.0%.³²

This sounding time is called phonation: it counts only the moments the folds actually vibrate, not the pauses between words.

The body notices quickly. In loud reading, an objective marker of vocal effort rises within the first 15 minutes, and people feel it after 30, in a study of 10 people.³³

After two hours of reading aloud at classroom volume, 86 teachers rated their voices mostly recovered within 4–8 hours, with a small remainder lasting a day or more. The paper disagrees with itself on the exact figures: its abstract gives 90% at 4–6 hours and full recovery at 12–18.³⁴

The closest existing job to talking to a computer all day is the call centre. Across 15 studies, 33–68% of operators had voice problems at some point in their career, and 27% at the moment they were examined.³⁵

One survey counted 6.8 voice symptoms per operator against 1.7 in the general population.³⁶

Direct evidence about dictation users comes only from the 1990s and 2000s, when dictation meant pausing after every word. In one series, five people who switched to speech recognition because of hand pain developed a strained voice; after voice therapy and shorter sessions, three of the five improved.³⁷

Muscle recordings in 10 people suggested the load moving rather than disappearing: less constant strain in the forearm and neck, and to a degree the shoulders, and a trend towards more in the cricothyroid, a muscle of the larynx.³⁸

Modern dictation removes the suspected cause — flat, machine-like speech with a pause after every word — and puts the microphone close to the mouth, so people can speak softly. That is a reasonable hope, not a finding.

No study has measured vocal load in people using today's dictation apps or the voice modes of AI models.

It is not the place. It is who is listening

Picture asking your phone for directions at home, then on a bus, then at your desk with colleagues around. For most people the answer changes three times.

The best numbers on this come from a small British experiment. 20 people aged 23–32 watched short videos of a man searching by voice or by text against a plain background, then ticked the places and the audiences where they would do the same. Voice beat text only behind the wheel, where the hands are busy.³⁹

Who was listening mattered more than where. Alone, voice was acceptable in 96% of answers; with friends in 83%, with a partner 78%, with family 71%. With colleagues, 26%. With strangers, 16%.

Among audiences, only colleagues and strangers made voice significantly less acceptable than text; among places, home, the passenger seat and work did. Try it on yourself before you look at their answers.

In a survey of 860 assistant users, shame lowered the intention to keep using one specifically in public, while pride raised it everywhere.⁴⁰

The norm moved anyway. In American surveys, using voice in a restaurant with friends rose from about 25% in 2017 to more than 40% in 2019. Men were 1.54 times as likely as women to use voice commands, and 46% said they were annoyed when others did it.⁴¹

Novelty wears off first. The shyness wears off too

Anyone who has bought a smart speaker knows the first week: timers, jokes, music, questions asked for the fun of it. Then the novelty wears off, and what is left is the habit.

How people change after they start is measured almost only on smart speakers at home. In a three-week study, 7 older adults were given an Echo Dot. Commands per week fell by a third: 59.6, then 50.1, then 39.7.

Music dropped from 23.9% of commands to 2.9%, while weather and shopping lists grew. One participant said she first “felt dumb talking to this machine”; by the end, all seven said they felt more confident and would keep using it.⁴²

Longer logs look different. In the Google Home histories of 18 families, averaging 58 weeks, use stayed fairly stable. Three families stopped using the speaker for weeks soon after buying it, over privacy, uselessness or a habit that had not formed, and later came back.

Families who quit for good were not studied. Control of the home grew as families bought more smart devices.⁴³

Where people stepped away, it was over usefulness, reliability or privacy, not because talking at home felt awkward.

A review of 69 papers on social acceptability points out that acceptability changes as people get used to an interaction, yet only three of them followed people trying an interface in everyday life; most asked people to imagine it, or tried it once in a lab.⁴⁴

That is the central gap for this page. I found no study that follows people who switched to dictation at work and measures, week by week, how their embarrassment changes.

Culture decides how loud voice is allowed to be

Whether you answer a message with a voice note or with text depends a lot on where you live.

In a 2023 YouGov survey across 17 markets, 66% preferred to send messages as text and 7% as audio. The preference for text was strongest in Britain (83%), Denmark (80%) and Sweden (78%) and weakest in the United Arab Emirates (38%); in Spain, 54% preferred text and 13% audio.⁴⁵

The mobile phone went through the same thing twenty years ago. In a 2007 comparison of students from five cultures at one university in Hawaii, talking on the phone on a bus scored 2.53 out of 5 among students from Japan, against 3.50 from Sweden and 3.44 from the mainland US.⁴⁶

Japan settled on a clear rule — no voice on the train, email is fine — and moved away from public voice as soon as a silent alternative existed.

In the US, 82% were irritated at least occasionally by loud phone calls in public in 2006, and the share who often ran into them fell from 50% to 39% by 2012.⁴⁷

By 2014, using a phone was fine on the street for 77% of Americans, on public transport for 75%, in a restaurant for 38% — and in a meeting for 5%.⁴⁸ It came from both sides: bystanders grew tolerant, callers grew quieter. And it stopped at the meeting-room door.

The office is learning to whisper

In 2026 AI dictation reached offices, and the first reports are stories, not numbers. Bloomberg describes it spreading from one desk to the next, with gooseneck microphones appearing and the whispering starting.⁴⁹

An analyst quoted by Computerworld calls embarrassment and the fear of distracting colleagues a major behavioural barrier. Thumbtack, a fully remote company, gave Wispr Flow to more than 200 staff, and its technology chief admits that getting used to speaking aloud, even alone in a room, took time.⁵⁰

The market is not waiting for the norm to change. It is making voice quieter. Wispr Flow sells a whisper mode and recommends speaking about a centimetre from the microphone.⁵¹

Research went further back in 2018: Microsoft's SilentVoice recognised short commands spoken while breathing in, leaking under 39 decibels of sound, with 1.8% of words wrong for a trained speaker and 7% for a new one.⁵²

People will trade accuracy for silence. In a 2021 experiment, 12 people were as willing to use silent speech — mouthing words without sound — with 10–15% errors deliberately injected as with none, while their willingness to speak aloud fell. All 12 put it down to privacy and security.⁵³

The fastest progress is not people growing braver. It is voice growing quieter.

The one place it is measured at scale: doctors' notes

Medicine is the only field where voice input backed by AI models has been measured on thousands of people. Ambient AI scribes listen to the appointment and draft the note.

In the largest study, published in JAMA in April 2026, about 1,800 users compared with 6,770 non-users saved 16 minutes of documentation per 8 hours of appointments, about 10%, with no significant change in after-hours work. Only 32% used the scribe in more than half of their visits.⁵⁴

In a randomised trial with 238 doctors, one scribe cut time per note by 41 seconds against 18 seconds in the control group, and another made no significant difference.⁵⁵ Burnout fell from 51.9% to 38.8% in 30 days, in a study without a control group.⁵⁶

Notes checked by hand had omissions in 18% of cases and invented details in 11.5%; 5.3% had an error that could cause serious harm.⁵⁷ Voice saved attention more than minutes, and the price is checking the text.

The experiments nobody has run yet

These are not small holes. Each card below is a question this page wanted to answer with a number and could not, because no study has measured it.

• A survey that counts dictation

  No survey with a random sample measures dictation or voice chat with AI as its own behaviour. The growth curve rests on company claims.

• The vocal dose of dictation

  Nobody has measured phonation time or vocal strain in people who dictate for hours with today's apps.

• A safe amount per day

  There is no evidence-based limit on how many hours of speaking a day are safe for the voice.

• Hands after switching

  No randomised trial shows that switching to voice prevents or treats strain injuries of the hands.

• Composing at a desk

  No controlled comparison of writing by voice and on a physical keyboard with modern recognition. The best data are from 1983 and 1999.

• A day spent talking

  No diary or long-term study follows people who work mainly by voice and measures their mental fatigue.

• Losing the shyness at work

  No field study follows new dictation users in an office week by week. Every long-term study of habit is about speakers at home.

How this was checked, and what I left out

I build Speak-Y, an app for voice input and meeting recording, and there is a banner for it at the bottom of this page. That is a reason to read this study with suspicion.

So every number that comes from a company about itself — including dictation companies that compete with mine — is marked in the text and drawn hatched in the charts.

The research was done in September 2026 along four questions: adoption, speed and accuracy, fatigue and thinking, social acceptability. Every number was traced to its original source where that was possible.

Small samples are kept, because they are what exists, but the sample size is always given next to the number.

My own reading of the whole picture, and it is an opinion: dictation will become normal at home, in the car and in a room of one's own, stay contested in the open-plan office and stay absent from meetings, just as the phone call did. What would change that is not braver people but quieter voice.