You use speech to text more than you might think. It writes the caption under a video, turns a voicemail into a text, and types the message you speak into your phone while walking. And for a growing number of people, it has become the main way they write.
This speech to text guide explains what the technology is, how modern systems work in plain words, the real difference between on-device and cloud processing, and how to choose a tool. It's a starting point: where a topic deserves its own page, we link to it.
A note on who's writing: I make LumeVoice, an AI dictation app. I've tried to keep this guide useful whichever tool you end up with.
What speech to text is#
Speech to text (also called speech recognition, automatic speech recognition or ASR) is software that converts spoken language into written text. It shows up in three main forms:
- Live dictation. You speak and the words appear in the app you're using: an email, a document, a chat box.
- File transcription. You give it a recording, such as an interview or a lecture, and get a transcript back.
- Voice interfaces. Assistants and voice modes turn your speech into text first, then act on it or answer.
The same core technology sits under all three. What differs is what happens to the text afterwards.
How modern speech recognition works, in plain words#
Researchers have worked on this for a long time. One of the earliest published systems, from Bell Labs in 1952, recognized spoken digits from a single speaker (Davis, Biddulph and Balashek, "Automatic Recognition of Spoken Digits", Journal of the Acoustical Society of America). For decades after that, systems were built from separate parts: one model for sounds, one for pronunciations, one for likely word sequences.
Today's leading systems are mostly single neural networks trained end to end. A well-documented example is OpenAI's Whisper, which its authors trained on 680,000 hours of audio paired with transcripts, across many languages (Radford et al., 2022). Here's roughly what happens when you speak to a system like that:
- Your microphone captures sound as a stream of numbers describing air pressure over time.
- The audio is turned into a picture of sound. The software converts short slices of audio into a spectrogram, a map of which frequencies are loud at each moment. Speech has recognizable shapes in that map.
- The model predicts words. The neural network reads those slices and predicts the most likely text, one small piece at a time, using everything it has heard so far as context.
- The text is formatted. Many models output punctuation and capitals directly; others add them in a separate step.
- Optionally, a language model cleans it up. AI dictation apps run a second model over the text to remove "um" and "uh", apply your corrections and fix grammar.
The important word is predicts. The model doesn't look anything up; it makes a best guess from sound and context. That's why it's very good at everyday language and less reliable with rare names, new jargon, heavy background noise or a microphone across the room. If you want to go deeper on what drives accuracy, see AI dictation accuracy: what affects it and how to test. For the step-by-step process when you convert a recording, see how audio to text converters work.
Transcription vs cleanup#
Step 5 is where tools differ most. Plain speech recognition writes what you said, including false starts. Cleanup turns "so the meeting is Tuesday, no, Wednesday" into "The meeting is Wednesday." That's useful for your own writing and wrong for anything that must be word for word, such as a quote. If you're unsure where dictation ends and AI writing begins, speech to text vs AI writing draws the line.
On-device vs cloud#
Every speech to text tool runs its model in one of two places, and this is the single biggest choice you'll make.
On your device. The model runs on your phone or computer. Your audio doesn't need to be uploaded, and it can work with no connection. The trade-off is that model size is limited by your hardware, so some apps offer a choice of smaller, faster models or larger, slower ones. Examples as of October 2026: Apple Dictation on iPhone in many languages (Apple), Gboard's advanced voice typing on Pixel phones, and apps such as Superwhisper and MacWhisper that run models on your computer.
In the cloud. Your audio is sent to a server, processed, and the text comes back. Servers can run large models quickly on any device, including an older laptop, and can add a cleanup step with a bigger language model. The trade-off is that you need a connection and you're trusting the provider with your audio, so read what they do with it: whether it's stored, for how long, and whether it's used for training. Windows voice typing (Microsoft), Wispr Flow and LumeVoice are cloud-based.
Neither is better in general. If your work involves confidential recordings or you often have no signal, on-device tools have a real advantage. If you want the same results on every device and heavy cleanup, cloud tools are usually stronger. Our explainer on hybrid voice architecture covers the middle ground.
Where speech to text is used#
- Everyday writing. Emails, messages, documents and notes. This is where dictation apps live.
- Prompts for AI assistants. Speaking a detailed request to ChatGPT, Claude or Gemini is often easier than typing it.
- Recordings. Interviews, lectures, meetings, voice memos and voice notes.
- Captions and subtitles for video and live events.
- Accessibility. For people with RSI, arthritis, dyslexia or low vision, speech to text can help reduce typing. See our accessibility page.
- Professional work. Clinicians, lawyers and court reporters have used dictation for decades, often with specialist tools.
- Developer products. Apps that add voice features through a speech API.
Speaking is also simply quick. A study by Stanford, the University of Washington and Baidu found speech input on a phone was about three times faster than typing on the phone's keyboard (Ruan et al., 2016). For what that means for real work, see voice dictation vs typing.
How to choose a speech to text tool#
Start with what you want to do, then work through these questions.
1. Live dictation or recordings? For writing as you go, you want a dictation app. For recordings, you want transcription software. Our roundups split the same way: best dictation software and best transcription software.
2. Which devices? Built-in tools are tied to one platform. If you switch between a Mac at work, a Windows PC at home and a phone, check that one tool covers them all. Platform roundups: Mac, iPhone, Android, and our talk to text apps list across all of them.
3. Does it work in every app? Some dictation only works inside one program, such as Word or Google Docs. System-wide tools type wherever your cursor is.
4. Raw text or cleaned-up text? If you dictate long passages, cleanup saves a lot of editing. If you need verbatim text, turn cleanup off or use a plain transcriber.
5. Your vocabulary. If you use names, product terms or jargon, look for a custom vocabulary feature.
6. Privacy and connectivity. Decide whether cloud processing is acceptable for what you'll say. If not, choose an on-device tool and check its language support.
7. Price. Try the free built-in option first. If you find yourself dictating every day and fighting its limits, a paid app is worth a trial; our affordable dictation apps roundup compares prices.
Start with what you already have#
As of October 2026, every major platform has free speech to text built in:
- Mac: Dictation, turned on in System Settings > Keyboard (Apple; our Mac dictation how-to).
- Windows: voice typing with Win + H (Microsoft; our Windows guide).
- iPhone: the microphone key on the keyboard.
- Android: the microphone in Gboard (Google).
Use one for a week. You'll quickly learn whether dictation suits you and what you'd want from a better tool.
Where LumeVoice fits#
LumeVoice is a cloud-based AI dictation app for Mac, Windows, iPhone and Android, with one plan for all four. On a computer you hold a key (fn on a Mac, Ctrl + Win on Windows), talk and let go; the text lands in whatever app has your cursor, with filler words removed, self-corrections applied, punctuation added and spoken lists formatted. It works in 100 languages and its desktop apps also transcribe audio and video files. Your audio isn't stored after processing and is never used to train models.
It isn't the right pick if you need to work without internet, use Linux or ChromeOS, or want a speech API for your own software. For those, the roundups above point to better options.
Try speech to text that writes like you meant it
Hold a key, say what you mean, and get clean text in any app on Mac and Windows, with iPhone and Android on the same plan. New accounts get 3,000 free words, no card required.
The short version#
- Speech to text predicts words from sound; it's excellent with everyday language and needs help with rare names.
- On-device tools keep audio at home and work without a connection; cloud tools run bigger models on any device.
- Pick by task first (live dictation or recordings), then devices, apps, cleanup, vocabulary and privacy.
- Try the free built-in tool before paying for anything.
