Browser tools

Speech tool

Speech to Text: Dictate Locally in Your Browser

Speak or choose an audio or video file and watch the transcript appear. English only. Everything is processed on this device.

Private · transcribed on this device

Dictate

Press the microphone and speak naturally. A short pause ends each sentence.

Audio source

Loading the page…

Mode: Live dictation · Change

Transcript

Click any sentence to correct it. Your edits stay put when new sentences arrive, and the transcript survives a reload.

Nothing transcribed yet

Press Start and speak, or choose an audio or video file. Sentences appear as the audio is processed.

Summary

Key points of what you say, updated as you speak.

Mode

Each mode downloads its speech model once. After that it loads from this browser, even offline.

Transcription mode

Why use an on-device speech to text tool?

Most dictation tools send your voice to a server to be transcribed. That is quick, but it means every word you say is recorded, uploaded and processed somewhere you cannot see. For a private note, a draft email or anything about work, that is a poor trade.

This page downloads an open speech recognition model once and then runs it inside your browser. Your microphone audio is turned into text on this device and is never sent anywhere, so it works the same on a train with no signal as it does at your desk.

Where the browser offers WebGPU it can also summarize while you talk: a small language model turns the transcript into a short list of key points and rewrites it as you go, so a long voice note, a meeting recap or a brainstorm ends with its summary already written. That runs on your device too.

What the tool helps with

  • Dictate notes, messages and drafts without typing
  • Transcribe a local audio or video file progressively, with timestamps from the recording and nothing uploaded
  • See your words appear while you speak, then corrected when you pause
  • See your voice's frequencies, pitch and microphone level in a live graph after Start
  • Fix any sentence in place before you copy the text
  • Get a live AI summary of the key points as you talk, written on your device with no account and nothing uploaded
  • Export timed subtitles as .srt or .vtt
  • Keep your voice private: nothing is uploaded, and it works offline once loaded

How to use it

  1. Pick a mode. Live dictation suits most computers; choose Light on a phone and Most accurate on a desktop with a good graphics chip.
  2. Press Start. The first time, the progress bar shows the model downloading; wait until it finishes.
  3. Allow the microphone when your browser asks, and speak normally.
  4. For a recording, drop an audio or video file anywhere on the page, or choose Audio / video file and pick one, then press Transcribe. Where the browser can decode the file itself it is transcribed several times faster than it plays. Pause or cancel at any time; completed sentences stay available.
  5. Pause briefly between sentences. Each one is written into the transcript when you stop.
  6. Click a sentence to fix a word, then copy the text or download it as plain text or subtitles.
  7. Open the Summary card and press Download and turn on once to see the key points of what you said, updated as you speak. Copy them with Copy summary.

FAQ

Can I transcribe an audio or video file from my computer?

Yes. Choose Audio / video file, drop or choose one recording, then press Transcribe file. Text appears sentence by sentence as the recording is processed. It follows the recording's timeline and pauses automatically if recognition needs to catch up. You can pause, resume or cancel; completed sentences stay. Timestamps and subtitles refer to the original recording. Starting a file creates a new transcript, so replacing existing text requires your explicit choice. The recording is read locally and is never uploaded or saved by this site.

Which recording formats can I use?

The tool uses your browser's audio and video decoder, so supported formats vary with the browser and operating system. The recording needs a playable audio track with English speech. If your browser cannot read it, try an audio-only copy in another format. Long recordings stream through the decoder rather than being loaded as one large decoded audio buffer. Keep this tab open while processing; a reload keeps the transcript, but you will need to choose the recording again.

What does the graph show while I speak?

After Start opens the microphone or a file starts processing, the graph shows recent sound: higher frequencies sit higher and brighter colours show sound above steady background noise. The pitch line follows voiced speech, the faint Level envelope shows microphone volume, and the green band marks where speech was heard. Level is relative to the microphone's digital range, not a calibrated loudness measurement. Stop keeps the last picture; the next Start clears it. With reduced motion, new sound fills the graph from left to right without scrolling. Nothing is recorded or uploaded.

Is my voice uploaded anywhere?

No. The speech model runs inside your browser, so your microphone audio is turned into text on this device and is never sent to this site or anyone else. The only thing downloaded is the model itself, and it comes from this site.

Why is there a download the first time?

Running the speech model on your device means your browser needs its own copy. Live dictation downloads about 190 MB, Light about 110 MB and Most accurate about 750 MB, once. The browser keeps them, so later visits start in seconds and work offline.

Which browsers does it work in?

Current versions of Chrome, Edge, Firefox and Safari, on computers and phones. Live dictation and Light use the graphics chip through WebGPU where the browser offers it and the processor otherwise. Most accurate needs WebGPU with half-precision support, which Chrome, Edge and Safari 26 provide on most Windows and Mac computers; on Linux it depends on the graphics driver. The live summary needs WebGPU too.

How accurate is it?

On clear English speech most sentences come out right. The usual mistakes are names, jargon and words that sound alike. Most accurate uses a model several times larger and makes fewer of them. Background noise and a distant microphone make every mode worse, and you can click any sentence to fix it.

Why do words change while I am speaking?

The grey line is a quick first guess, refreshed about once a second while you talk. When you pause, the whole sentence is transcribed again with all of its context and written into the transcript, which is usually more accurate.

Can it summarize what I say?

Yes, where the browser offers WebGPU. Open the Summary card and it lists the key points of your transcript and rewrites them as you speak, using Qwen3.5 0.8B, a small open language model that runs on your graphics chip. It is downloaded once from this site, about 570 MB, and runs on your device, so neither the transcript nor the summary is sent anywhere. A model this small can misread or leave out a point, so check the key points against the transcript before you rely on them.

Why does the Summary say it needs WebGPU?

The summary model runs on the graphics chip through WebGPU, which current Chrome and Edge offer on Windows, macOS and ChromeOS, as do Safari 26 and Chrome on many Android phones. Chrome on Linux offers WebGPU only with Vulkan turned on. On the processor alone the model would be far too slow to keep up with you, so without WebGPU the Summary card only explains what it needs. Dictation works the same without it.

Does it understand languages other than English?

Not yet. Every model here is set up for English, and speech in another language comes out as English-sounding guesses.

Can I make subtitles with it?

Yes. Turn on Timestamps to see when each sentence started, then download .srt or .vtt subtitles. The times come from when you started and stopped speaking. A long sentence split across several subtitles shares its time out by length, so those splits are approximate.

What happens to my transcript if I close the tab?

It is kept in this browser's storage, so it is still there when you come back, and it is never uploaded. Clear removes it. The audio itself is never stored.

How do I free the space the models use?

Press Remove downloaded models beside the modes, then press it again to confirm. The models download again the next time you press Start.

Models and credits

Live dictation and Light use Moonshine Base and Tiny by Useful Sensors, Most accurate adds OpenAI's Whisper Large v3 Turbo, and Silero VAD detects when you are speaking. All are open models released under the MIT licence and run here through Transformers.js and ONNX Runtime Web. Every file is served by this site at a pinned version.

The summary is written by Qwen3.5 0.8B from Alibaba's Qwen team, released under the Apache 2.0 licence, in the ONNX export by onnx-community. It runs on your graphics chip through WebGPU, so this site never sees the transcript or the summary.

All tools Back to home Privacy