The Future of Typing: Voice, AI and the Keyboard

Keyboards are not going away, and the change will be uneven. Voice is clearly faster for getting a first draft onto the screen, and AI is changing what people type, from the exact words to a description of the result they want. Editing, precise selection, shared offices and private moments keep the keyboard in use. Brain-computer interfaces (BCIs) now decode speech for people who cannot speak, and in one long-term home study the average rate was about 9 percent above the average of people in a large online typing test ¹, but those results come from a handful of people. For the short answer, read will voice replace keyboards. The rest of this guide explains each part and the evidence behind it.

The starting point is the speed gap. In a controlled study on an iPhone, English speech input reached 153 words per minute, against 52 on the built-in keyboard ^{2 3}. The phrases were short, read aloud, and spoken in a quiet room, so the figure shows what speech can do under good conditions. It does not show what a working day looks like.

Voice will take over composing, not editing

Voice wins when you are making new text and loses when you are changing text you already have. The phone study counts the time spent fixing errors while entering text, but it does not measure the time spent editing a dictated draft afterwards ². Editing is where the keyboard keeps its advantage.

A longitudinal study of continuous dictation points the same way. Navigation and error correction took a large share of the dictation task in the early stages of use ⁴. Recognition has also improved: the 2018 Ruan paper says upper-bound performance, meaning the best case, has greatly improved on earlier systems ⁵. The structure of the problem has not changed, though. Speaking a sentence is fast, and finding and changing a specific word in the middle of a paragraph is slow.

Voice assistants have tried to close this gap with editing commands. Android's Voice Access, for example, includes a tutorial that covers basic text editing alongside tapping and scrolling ⁶. Those commands work best for simple changes, while a keyboard and mouse remain more precise for selecting exact characters, cells or ranges.

Our reading of the evidence gives this split:

The pattern is that speech is the input for creating and the keyboard is the input for revising. The better tools put the two side by side instead of choosing one. For a fuller account of how dictation changes writing, see how to dictate and dictation mistakes.

Voice-first devices stalled partly because talking is public and slow to correct

Smart speakers were meant to make voice a normal way to use a home, and ownership has barely grown since. In the United States, ownership rose only from 33 percent in 2021 to 35 percent in 2025 ⁷, which the publisher calls a plateau. Many owners still value the device. In a three-month study of 39 adults aged 65 to 85 in subsidized housing, 72 percent wanted to keep using their smart speaker afterwards ⁸. That study is small, but it fits a pattern: the devices work best in private, everyday settings. The figures are set out in are smart speakers still popular.

The sources point to two reasons, though no single study measures how much each one matters. The first is that talking is public. A 2018 workshop paper from University College London expected that users would be more self-conscious speaking their answers than typing them discreetly ⁹. The authors built a speech prototype for public spaces and observed that people in a coffee shop avoided it, while people at a crowded event were more open to it. You can read the CHI 2018 workshop paper for their observations.

The second reason is that speech is slow to correct. A recognizer has to decide when you have finished speaking, and it can cut you off too early. In a 2023 study of people who stutter, 61 percent of those familiar with voice assistants named being cut off before they finished talking as a reason they used them less often ¹⁰. Other users run into the same problem when a pause in the middle of a sentence ends the recording, and a typed interruption does not have that risk.

Noise is the third failure. A 1997 review of recognition experiments reported that a machine with noise adaptation reached 12.8 percent error at a signal-to-noise ratio of 10 dB, more than ten times the 1 percent error of human listeners in the same comparison ¹¹. Modern systems do much better, and the review is old, but the pattern is the same: a room full of other sounds makes speech harder to recognize, and it also makes speaking in that room awkward. For the practical side, see dictation background noise.

AI shifts typing from writing words to describing outcomes

AI changes the unit of input. For decades, typing meant putting the exact words that would end up in a document onto the screen. With a language model, many people now type or say a description of the result they want, and the model writes the words. The guidance from AI vendors describes this shift directly. Anthropic's prompting guidance says to think of Claude as a capable new employee who lacks context, and that the more precisely you explain what you want, the better the result ¹². A prompt is therefore a briefing, and a briefing is often easier to speak than to type.

Vendors have built dictation into their assistants for this reason. Microsoft describes dictation in Copilot as a way to speak your prompt instead of typing it ¹³. OpenAI's help article for ChatGPT says to use dictation when you want to record a prompt, review and edit its transcription, and then send it as text ¹⁴. The second half of that advice matters. The speaking happens first, and the keyboard returns for the review. This is the main reason we expect typing to change its role, not disappear: people will describe, speak, then type to correct.

The same review step explains why AI does not remove the keyboard from writing. A model's output still has to be checked, and checking is a reading and editing task, which favors the keyboard. For the basics of prompts, see the AI prompt entry, and for how spoken prompts differ from typed ones, see speech to text for AI.

Silent speech interfaces read words you do not say aloud

Silent speech is the recognition of words from muscle signals or brain activity without any sound. It matters for people who cannot speak aloud, and for anyone who needs to dictate in a public place. There are two main approaches: sensors on the face or throat that record muscle activity, and implanted electrodes that record the brain.

The face-sensor approach has produced strong results in the lab. In the 2018 AlterEgo study, a wearable that reads silently mouthed words reached a median word accuracy of 92 percent in its user studies ¹⁵. In 2020, Gaddy and Klein trained a system on silent facial muscle signals and turned them into audible speech. In one data condition, the transcription word error rate fell from 64 percent to 4 percent, and in another it fell from 88 percent to 68 percent ¹⁶. The result depends heavily on the condition, and the test was a laboratory dataset, not everyday use.

The most precise silent speech result we found is from the implant study. In a copy task, the participant's silent speech was decoded at 49.7 words per minute with 96.5 percent accuracy ¹⁷. His vocalized speech in the same task reached 30.6 words per minute with accuracy above 99 percent ¹⁸. That makes the silent condition about 62 percent faster ¹⁹, at the cost of some accuracy. The gain is real, but it came from one person with an implant, not from a device anyone can buy.

So can you dictate without speaking aloud? The systems in these sources are research prototypes. Those that read facial muscles need sensors on the face, and the only silent speech results here that report a speed come from an implant. None of the sources we checked describes a consumer product that dictates from silent speech. For the plain answer, see what is silent speech.

Brain-computer interfaces now decode speech for people who cannot speak

A brain-computer interface records activity from the brain and maps it to words or sound. The clearest case for it is a person who cannot type or speak. In a 2023 study, the participant could not speak or make speech sounds, and could not type, because of weakness in her facial, vocal, arm and hand muscles ²⁰. Her attempted speech was decoded into text at a median of 78 words per minute, with a median word error rate of 25 percent ²¹.

The chart compares the main measured rates in this guide. The bars mix different measures and different people, so read the subtitle and the caveats before comparing them.

The studies differ in design, but the trend is consistent. Each reached a useful speed in a small group of people:

The 2024 report shows how he typed at the time. He used a gyroscopic head mouse, at 6.3 correct words per minute ²⁴. Decoding speech is a different route to text, not a faster head mouse.

The 2026 study is the longest. The participant used the device at home for more than 3,800 hours, with no researchers present ²⁵. He communicated at an average of 56 words per minute ²³. That average is about 9 percent above the 51.56 words per minute of participants in a large online typing test ^{1 26}. The two figures come from different kinds of use, so the gap is a rough comparison, not a head-to-head test. It is also about 2.7 times slower than English speech on a phone ²⁷.

Voice quality is the other test. A 2025 study turned neural activity into a synthesized voice, and the voice was often intelligible, but not consistently ²⁸. That is an important limit. A speech system that is sometimes unintelligible is not yet a substitute for talking in ordinary conversation.

The brain results here come from a few people with paralysis, and each used electrodes placed in the brain by surgery. The 2026 study itself reports one participant with severe dysarthria ²⁹. These are real achievements for people who need them, but they describe what is possible with an implant, not what most people will use. For the direct answer on speeds, see how fast can a brain-computer interface type.

Most people will mix voice, keyboard and AI for years

The likely future is a mix, not a replacement. Voice is the faster way to get a rough draft down, the keyboard is the way to revise and to work in public or shared spaces, and AI sits between them, turning a description into text that a person then checks. None of the measured rates in this guide suggests that voice or a brain interface will make typing unnecessary for most people in the near term. The fastest typists in the online test reached 120 words per minute or more ³⁰, which no measured brain interface or silent speech system in this guide has matched.

Here is how we would split the work today:

For writers who dictate on a Mac or Windows PC, a dictation app that runs on the device handles the drafting step, and the keyboard handles the edits. Yap, which we make, is one such app. The built-in options are often enough for short messages, and you can compare them in how to choose a dictation app.

Speech is also not uniformly fast or reliable for every speaker. Recognition accuracy varies by speaker and by setting, and noise is one of the sharpest drops, as the 1997 figures above show. That is why the guides on speech recognition for stuttering and dysarthria and dictation background noise matter as much as the headline speeds.

Two habits follow from the evidence. First, keep your keyboard skills sharp, because the review and editing steps still depend on them. A dictated draft that you cannot correct quickly can cost more time than it saves. Second, decide in advance where you will not speak. A shared office, a train or a client call is a poor place to dictate even when the recognizer is accurate, because the discomfort the 2018 workshop describes is about being overheard, and accuracy does not fix that. Neither habit needs a new product. It needs a clear split between drafting by voice and revising by keyboard.

What to watch through 2027 is narrower than the hype suggests. The useful signals are whether the home BCI results reach more than one person, whether silent speech moves from research prototypes to something people can buy, and whether speech systems become reliable in public settings. Each would change the numbers in this guide. All figures here are as checked on 2026-10-17, and the sources are linked from each fact. For how voice assistants and dictation differ in practice, read voice assistants vs dictation, and for where AI agents are heading, read AI voice agents. For the wider subject, go back to the voice and AI overview.