Live Transcription (beta)

Near-real-time speech-to-text, entirely on your device — experimental.

Transcribe speech to text live as you talk, in real time, with no upload. Free real-time transcription that runs in your browser — nothing is recorded or sent.

About Live Transcription (beta)

Live Transcription (beta) is a free, experimental browser tool that turns your microphone speech into text in near-real time, entirely on your device. It captures mic audio, downsamples it to a rolling 16 kHz window, and runs OpenAI's Whisper model (Transformers.js) locally — using WebGPU when available and falling back to single-threaded WASM — so audio is never uploaded and there is no signup. Use it for quick on-device dictation, captions, or note-taking, and to test how well live speech-to-text performs on your specific hardware; accuracy and latency vary by device.

How to use Live Transcription (beta)

  1. On desktop, pick a model: Tiny (fast, ~40 MB, English-only) or Base (more accurate, ~145 MB). On mobile only Tiny is offered, since Base is too heavy for a phone.
  2. Click 'Start live transcription' and allow microphone access when the browser prompts.
  3. Wait for the speech model to download once (you'll see a 'Loading model… %' progress); after the first run it's cached and works offline.
  4. Speak naturally — text appears as you talk, with in-progress words shown in grey and finished phrases locking in as you pause.
  5. Watch the HUD (Backend, Model, Model load, Last inference) to gauge whether your device keeps up live — if 'Last inference' stays well under 1.3s, it's running in real time.
  6. Click 'Stop' to end; a final pass runs over the buffer for the cleanest transcript, then you can copy the text out.

Frequently asked questions

Is my microphone audio uploaded to a server?
No. Your mic audio is captured and transcribed entirely on your device using a local Whisper model — nothing is ever uploaded. Only the model files and runtime download once from public CDNs, after which it works offline.
Is it free and do I need an account?
Yes, it's completely free with no signup, account, or watermark — everything runs locally in your browser.
Which models can I use, and what's the difference?
Two are offered: Tiny.en (~40 MB), an English-only model that's fastest for live use, and Base (~145 MB), which is more accurate. Base is desktop-only because it's too memory-heavy for a phone, so mobile is locked to Tiny.
Why is this labeled experimental / beta?
It's a prototype built to measure whether a small on-device Whisper can do near-real-time transcription on real phones and desktops. Accuracy and latency depend heavily on your hardware, so it bypasses the desktop-only gate used by the polished tools.
Does it need a fast GPU, and what does the HUD show?
It uses WebGPU when your browser supports it and otherwise falls back to single-threaded WASM (q8). The HUD reports the active backend, model, one-time model-load time, and last inference time so you can see if your device transcribes faster than each ~1.3s tick.
How does the text stay readable instead of rewording itself?
It transcribes in roughly 12-second segments and commits a phrase when you pause for about 0.7s after real speech. Each finished utterance locks in and appends, so the transcript builds cleanly rather than overwriting the growing window every tick.

People also search for

Live Transcription (beta) is also known as live transcribe app, real-time speech to text, live captions generator, voice to text online, browser dictation tool, on-device transcription.