Speech to Text AI: Dictation, Assistants and Voice Agents

Speech to text is AI. Every mainstream engine of the last decade is a neural network that learned to recognize speech from recorded examples, so it counts as machine learning in the technical sense. It is not generative AI by default, because a plain recognizer transcribes what it heard and does not write new text. The two meet when a large language model edits the transcript afterwards: removes the ums, adds punctuation, applies a spoken correction, matches the tone of the app you are typing in. That second step is optional, and it is what the new generation of AI dictation apps sells.

Three things follow, and the rest of this page is the evidence for them.

1. Dictation and assistants do different jobs. Dictation puts your own words on the page. An assistant or a voice agent interprets what you mean and does something. Pick by the job.
2. The AI edit is a trade. It saves cleanup time, and it costs cloud processing, a delay after you stop talking, and a small chance that your words come back changed.
3. Speaking is still the faster way to compose. A controlled study found speech about three times faster than a phone keyboard ¹. AI changes what happens to the words afterwards, not that advantage.

If you want to know how to use dictation rather than how it relates to AI, start with our overview of dictation. If you want the recognizer itself, see our overview of speech recognition.

AI dictation is speech recognition plus an optional language-model edit

The pipeline has two stages. A recognizer turns audio into words. In the newer apps, a language model then turns those words into finished text. Our guide to AI dictation draws the whole chain, with the delay each step adds.

The recognizer is the part you cannot do without. OpenAI's Whisper, one of the best-known open models, was trained on 680,000 hours of multilingual audio with transcripts ². Apple, Google and Microsoft built their own recognizers on the same principle, and all three now ship neural models in their built-in dictation. We cover which is which in the guide to whether speech to text is AI and in the question is speech to text generative AI?.

The second stage is where products differ. Microsoft says Fluid dictation in voice typing on Copilot+ PCs automatically corrects grammar, punctuation and filler words as you speak ³. Wispr Flow says it removes filler words automatically ⁴. Apple says that with Apple Intelligence in iOS 27, iPhone dictation uses an on-device model to improve spelling, punctuation and capitalization, in English on the iPhone 17 Pro, iPhone 17 Pro Max, iPhone Air and later models ⁵. So the edit step is arriving inside the built-in tools, and some of it runs on the device rather than in the cloud.

What the edit costs depends on where it runs. Wispr Flow's privacy page says dictated audio is sent to its servers over an encrypted connection, so cloud processing is part of that design ⁶. The same company says it deletes audio as soon as it is transcribed unless you turn on cloud storage ⁷. An open-source app such as HEX describes the opposite default: ordinary dictation uses a local model, and only when you turn on post-processing does it send the transcript, the foreground application and the hostname of the Brave website to the model provider you configured ⁸. Both are reasonable designs. They are different bargains, and you should know which one you have chosen. For the privacy side in depth, see our overview of voice privacy.

Common questions on this part:

• Is AI dictation more accurate than built-in dictation?
• Why is AI dictation slower than built-in dictation?
• Does AI dictation read your screen for context?
• Which AI models do dictation apps use?
• Can AI dictation run entirely on your own computer?

If you are choosing an app, our overview of dictation software compares them. Free tools are often enough. Apple's, Microsoft's and Google's built-in dictation need no subscription, and our page on online dictation works in the browser for free.

Whether speech to text counts as AI depends on the question you are asking

The technical answer is yes. The practical answer depends on the rule you care about. Our guide to whether speech to text is AI gives the answer per product from each vendor's own documentation, and it separates three different meanings of "AI".

The last row matters in schools and workplaces. A rule that bans "generative AI" usually does not reach a recognizer that only transcribes, and a rule that bans "AI tools" without definition may. The wording of the policy decides, not the technology. We take this up in the question does voice typing count as using AI? and, for the writing side, in is dictated text still your own writing if AI cleans it up?.

The history matters less than the cut-off. Recognition used statistical machine learning long before deep learning, and the shift to neural networks came in the 2010s. Our overview of speech recognition gives the timeline. The questions people ask most about each vendor are answered in does Microsoft dictation use AI? and does Google voice typing use AI?.

There is also the opposite request: dictation with no AI rewrite. It is possible, and it is mostly a setting. The guide to speech to text without AI shows which tools type exactly what you said and how to switch AI cleanup off. Two points from it are worth stating here. First, every recognizer uses machine learning, so "without AI" can only mean without generation, without a cloud service, or without training on your recordings, and each of those is a separate choice. Second, on-device recognition keeps your audio away from cloud AI services entirely. Apple says that transcription with its SpeechAnalyzer developer API is entirely on the device, although the language models must first be downloaded ⁹.

Related questions: how do you turn off AI rewriting in a dictation app? and can AI detectors flag dictated text?. For definitions see AI text detector.

AI cleanup removes fillers and fixes text, and it can also change what you said

The case for cleanup is simple. Spoken language is full of fillers, false starts and repairs, and written text is not. A 2006 review by Lease and Johnson reports that fillers and speech repairs may be deleted from transcripts without compromising meaning ¹⁰. A language model does this deletion, and more: it can apply a spoken correction ("Tuesday, no, Thursday"), insert paragraphs, and format a list. Our guide to AI cleanup for dictation shows the levels from verbatim to full rewrite with before-and-after examples.

The same power is the risk. A model that rewrites can drop a qualifier, flatten your voice into generic prose, or add a word you never said. The nearest measured example comes from recognition itself. In a 2024 study, roughly 1 percent of Whisper audio transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio ¹¹, and 38 percent of those hallucinations included explicit harms such as inventing associations or implying false authority ¹². That is a recognizer, not a cleanup model, and the rate is small. But it shows that a system built to produce fluent text will sometimes produce fluent text that nobody said.

A related finding concerns writing with language models in general. In an experiment on writing with an opinionated language model, using the model affected the opinions people expressed in their writing, and shifted their opinions in a later survey ¹³. Cleanup is not the same as suggestion, but the lesson carries over: an editor you do not watch can move your meaning.

Three habits keep the edit faithful.

1. Choose the lightest level that works. Fillers and punctuation rarely change meaning. Full rewrites often do.
2. Review what matters. Names, numbers, dates and anything legal or medical deserve a second look whatever the tool.
3. Keep the raw transcript. If the result looks wrong, you can compare.

You can also run cleanup yourself with any chatbot and a fixed prompt, which keeps the transcript local until you choose to paste it. The question how do you clean up dictated text with ChatGPT? has a prompt that works. More questions: does AI dictation change what you said?, how does AI remove filler words? and does AI cleanup make writing sound generic?.

Language models change dictation by correcting, transcribing and prompting

Large language models enter speech recognition in three ways, and our guide to speech to text and LLMs separates them.

Correcting the recognizer's output. Every recognizer already contains a language model that scores which word sequences are plausible. An external LLM can do this job on a far larger scale. The HyPoradise benchmark was built for exactly this test: it holds more than 334,000 pairs of the recognizer's candidate transcripts and the correct text ¹⁴. Its authors report that, because an LLM generates text rather than only choosing among candidates, it can correct words that were missing from every candidate ¹⁵. This is called generative error correction. It can help on hard audio, and the same generation is how invented words can enter.

Transcribing directly. Speech language models take audio in and produce text, with no separate recognizer. OpenAI's price list shows its gpt-transcribe model at $0.0045 per minute ¹⁶. Whisper itself is not a large language model in the chatbot sense, and our question is Whisper a large language model? explains why.

Taking spoken input as prompts. When you speak to a chatbot, the speech becomes a prompt. This is the subject of the next section.

Whether an LLM belongs inside a live dictation pipeline is a question of delay and cost, not only accuracy. A language-model pass has to wait for the words, and every added step is time between your last word and the text appearing. In a 2025 test, Apple's speech API transcribed a 7-minute-31-second episode in 9 seconds ¹⁷, while Whisper Large V3 Turbo took 40 seconds ¹⁸, but Whisper's word error rate was 1.5 percent against Apple's 10.3 percent ^{19 20}. The author cautions that exact numbers would differ under slightly different methods ²¹. The lesson is the trade itself: the slower, larger model made far fewer errors on that episode. Which end of the trade you want depends on whether you are dictating a quick message or finishing a document.

Related questions: do LLMs make speech recognition more accurate? and which speech to text should you pair with an LLM?.

Speaking a prompt is faster than typing it and gives the model more to work with

The first reason is speed. In a controlled iPhone study, English text entered by speech reached 153 words per minute against 52 on the built-in keyboard ^{1 22}. At those speeds a 1,000-word draft takes about 6.5 minutes spoken and 19.2 minutes typed, a saving of about 12.7 minutes ²³. The study also found that speech produced fewer errors during entry than the phone keyboard ²⁴, though a laboratory with short phrases in a quiet room is not an office.

The second reason is content. When prompts cost effort, people write short ones, and short prompts leave the model guessing. When speaking is easy, people include the background, the constraints and an example. That is a claim about behavior that our guide to speech to text for AI examines with the available research, and the question do longer prompts get better AI answers? states what is and is not established. We do not claim that longer is always better. We claim that a spoken prompt is more likely to contain the context a model needs.

Models also read through most of the noise. Fillers and the occasional misrecognized word rarely stop a chatbot understanding the request, because the model infers from context, the same mechanism that makes error correction work. See do AI chatbots understand prompts with dictation errors?. Where voice prompts still break is on code, unusual names and symbols, where a wrong word changes meaning and the context cannot repair it.

A spoken prompt works best with four parts: the goal, the context, the constraints and the output you want. The question how do you dictate a good prompt for ChatGPT or Claude? gives examples. See also prompt (AI).

Two modes are easy to confuse. In Claude, dictation converts your speech to text so you can type prompts by speaking ²⁵. Voice mode is a separate feature for talking with Claude. Use the chatbot's own microphone for short prompts, and system-wide dictation for long ones that you want to review before sending. For coding agents, our page on dictation for Claude Code covers the setup.

Voice notes become structured text after the transcript exists

Recording a rambling thought and getting back a summary or a to-do list is the other big use of AI beside dictation. Our guide to AI voice notes breaks it into steps: transcribe, then structure, then review. The transcript step is a recognition problem, covered in our overview of transcription. The structuring step is a language-model job, and a summary, a task list and a draft are three different jobs for it.

Phones and note apps now do the structuring themselves, some on the device and most in the cloud. The question can your phone summarize a voice recording on its own? lists what Pixel, iPhone and Samsung phones do. If you use a chatbot instead, can ChatGPT summarize a voice memo? answers it, and how do you turn voice notes into a to-do list? gives a workflow.

The caution is the same as for cleanup: a summary is a rewrite, and a summary leaves things out and sometimes adds them. Keep the transcript. Check names, numbers and decisions against it. The question are AI summaries of voice recordings accurate? covers the evidence. For a clinical example of measured accuracy, a 2025 neurosurgery pilot found a Whisper-based model had a mean word error rate of 1.75 percent counting all errors ²⁶, and Dragon Medical One's rate was 2.7 times as high once linguistic errors were excluded ²⁷. That is a transcript measure, not a summary measure, and a small pilot, but it shows how close a modern recognizer can get when the audio is good.

Voice assistants act on your intent, and dictation records your words

A voice assistant is built to understand what you want and do it: set a timer, answer a question, send a message. Dictation is built to record exactly what you say. That difference explains most of the frustration people have when they try to dictate to an assistant. An assistant decides when you are done and what you meant. Dictation waits and writes. Our guide to voice assistants vs dictation sets out the difference. For the day-to-day Apple comparison, see dictation vs Voice Control vs Siri on a Mac.

Between 2024 and 2026 the assistants turned into language-model chatbots, which blurs the line without erasing it. The facts from the vendors, with dates:

Two details need care. The Google announcement is a plan reported in March 2025, and rollout dates and exceptions have moved since, so treat our summary as a dated snapshot and check Google's help pages for your device. Amazon's price is a vendor price on the day we read it.

Because these assistants are now chatbots, the old rule of thumb still holds with an update. Use an assistant for actions and answers. Use dictation for anything you will send or keep, because dictation does not decide when you are finished and does not paraphrase you. Our guide explains why assistants are poor at long text, and the question can Siri, Alexa or Gemini take long dictation? answers it directly. Other questions: which is the smartest AI voice assistant?, what is Alexa+?, is the new Siri powered by Google Gemini? and why do voice assistants misunderstand you?. The terms voice assistant, wake word and intent recognition are defined in the glossary.

Google Assistant commands and the move to Gemini

If you use Android, the practical question is which commands still work. Our guide to Google Assistant voice commands lists them by category (calls, messages, timers, smart home, dictation), marks the ones that changed in the move to Gemini, and covers phones, speakers, cars and watches, which move on their own schedules. Common questions: is Gemini replacing Google Assistant?, can you switch back to Google Assistant?, how do you turn on Hey Google? and which Google Assistant features does Gemini lack?.

A home voice assistant can stay entirely in your house

Not every assistant lives in someone else's cloud. Home Assistant's Assist runs a pipeline of wake word, speech to text, an intent engine (or an optional language model) and a spoken reply. Our guide to Home Assistant voice walks through each stage. The project documents two local speech-to-text options with a clear trade. Speech-to-Phrase is a closed-ended model that transcribes what it knows, in under one second even on a Raspberry Pi 4 ³¹. Whisper is open-ended and tries to transcribe everything, but takes around 8 seconds per command on a Raspberry Pi 4 and under a second on an Intel NUC ³². The same fixed-versus-open trade runs through all of voice AI: a model that listens only for a known list is fast and cheap, and a model that listens for anything needs more power.

Questions: what is the best speech to text for Home Assistant?, what voice commands does Home Assistant understand?, what is the Home Assistant Voice Preview Edition? and can Home Assistant use ChatGPT or a local LLM for voice?.

Voice agents are built on delay, and people notice a gap of a fraction of a second

A voice agent listens, decides and speaks back in real time. It is the assistant idea applied to phone calls and customer service, usually without a screen. Our guide to AI voice agents explains how they work and where they fail. The definition is in the glossary under voice agent.

Two designs exist. A cascade chains three models: speech to text, a language model, then text to speech. A speech-to-speech model does the whole step in one network. The cascade is easier to control, log and fix, because you can read the text between stages. The single model removes the hand-offs and with them some delay, but it is harder to constrain. The question what is the difference between cascaded and speech-to-speech voice AI? goes through it.

The benchmark is human. Across ten languages, the mean gap between one speaker finishing and the next starting in conversation was 208 milliseconds ³³. A person answers almost as the other stops, often before they have finished planning the whole answer. An agent that needs a full second to reply therefore feels slow even when the answer is good. The hard part is not only speed. It is deciding when you have finished talking. A pause for thought is not the end of a turn, and a system that cuts in too early is as irritating as one that waits too long. This is the job of turn detection. See how fast does a voice AI agent need to respond?.

Agents fail in predictable places: names, numbers and addresses, background noise and interruptions. Whisper Large V2 had a word error rate of 25.5 percent on CHiME6, a test set of distant, noisy, conversational speech ³⁴, about 9.4 times its error rate on clean LibriSpeech audio ³⁵. A phone line in a noisy room is closer to the first case than the second.

Law applies once an agent calls a person. In February 2024 the US Federal Communications Commission ruled unanimously that AI-generated voices are "artificial" voices under the Telephone Consumer Protection Act ³⁶. That brings AI callers under the rules on artificial and prerecorded voices, which generally require the recipient's consent. Other countries and states have their own rules and some require disclosure. This is a summary of one ruling and not legal advice. Our guide covers the disclosure question, and the question do AI voice agents have to disclose that they are AI? is the place to start. Cost is covered in how much does an AI voice agent cost per minute?.

Cloud speech costs fractions of a cent a minute, so the real cost is everything around it

Speech recognition in the cloud is cheap per minute and adds up per user. OpenAI lists gpt-transcribe at $0.0045 per minute ¹⁶. At that price an hour of audio costs about $0.27 through the API, against about $119.40 at Rev's $1.99 per minute for human transcription ³⁷, roughly 440 times more ³⁸. A language-model cleanup pass adds a per-word charge on top. Our guide to AI dictation cost works out the per-user figures from published prices and the volumes a daily habit produces.

A heavy user changes the picture. Because speech is about three times faster than typing, someone who drafts by voice produces far more words per hour, so ten hours of keyboard drafting becomes about 3.4 hours of speaking ³⁹ and the cloud bill scales with the hours spoken. That is why cloud AI dictation is normally sold as a subscription: the provider carries a variable cost for every minute you use. Pricing varies by app. Wispr Flow, for instance, lists a free desktop allowance of 2,000 words per week ⁴⁰. These are vendor plans on the day we checked, and they change.

On-device dictation has no per-minute cost, because the work runs on a computer you already own. What it uses instead is memory, battery and storage. Apple says its improved on-device Dictation model in macOS 27 is available on Macs with M3 or later and at least 12GB of unified memory ⁴¹. On-device is not free in the sense of costing nothing. It moves the cost onto hardware and removes the subscription and the cloud copy of your audio. Yap, which we make, takes the on-device approach on Mac and Windows; see our pricing page for what that costs and offline dictation on Mac for how it works. The comparison below is fair to both.

Questions: why do AI dictation apps charge a monthly subscription?, is on-device dictation cheaper than cloud dictation? and does on-device AI dictation drain your battery?.

Voice will take over composing, not editing

Will voice replace the keyboard? The evidence points to a split. Voice wins at composing and at short commands, because speaking is faster than typing and needs no hands. It loses at editing, precise selection, shared spaces and private moments. The study behind the speed numbers above counted corrected and left-over errors: speech made fewer errors during entry yet left slightly more in the final text ²⁴, which fits the pattern that voice gets text down quickly and a person then fixes it. A separate 2022 study found that speech recognition was faster than typing on mobile, but the error rate for typing was lower ⁴².

Our guide to the future of typing weighs the rest of the evidence: smart speakers, which reached many homes and then plateaued as devices for timers and music rather than writing, the places voice interfaces fail, and what comes next. The questions will voice replace keyboards? and are smart speakers still popular? give the details.

Two further horizons are real but far from your desk. Silent speech interfaces read words you do not say aloud, and the question what is silent speech, and can you dictate without speaking aloud? explains the state of that work. Brain-computer interfaces decode attempted speech for people who cannot speak. In one 2023 study a participant's attempted speech was decoded at 62 words per minute ⁴³. In another, attempted silent speech was decoded at a median of 78 words per minute with a median word error rate of 25 percent ⁴⁴. These are single participants with implanted devices, and word errors remain frequent. They matter as proof that speech can be read from the brain, and they matter most to people with paralysis. They are not a product you can buy. See how fast can a brain-computer interface type?. Our overview of accessibility covers the tools in use today.

AI changes what people type more than how. Increasingly the input is a description of an outcome ("summarize this thread and draft a reply") instead of the words themselves, and that kind of input suits the voice. The likely result for years to come is a mix of voice, keyboard and AI, and not a single winner.

How to choose: the job decides

The decision is shorter than the topic suggests.

1. You want your own words on the page. Use dictation. Start with the free built-in tool on your device (see our guides for Mac, Windows and phones and tablets).
2. You want those words polished. Add AI cleanup, at the lightest level that works, and review names and numbers. Know whether your audio goes to a cloud service.
3. You want the exact words, no AI. Use verbatim dictation and switch off cleanup. The guide to speech to text without AI shows how.
4. You want a task done. Use an assistant. For long text it is the wrong tool.
5. You want a system to answer calls. That is a voice agent, with its delay budget and its legal rules.
6. You are briefing an AI. Speak the prompt: goal, context, constraints, output. See speech to text for AI.

For how typing and speaking compare in raw speed, see our overview of typing speed, and for the research on writing by voice see dictation. If privacy decides it, see voice privacy. If you want to try dictation now, the download page has the app and /online-dictation works in the browser with nothing to install.

All the guides on voice and AI