DevToolPal
EN

Live Transcription

Dictate in English and watch your words appear as text, transcribed live in your browser by the Moonshine speech model. No audio is uploaded.

Runs locally, never uploaded

Options

Model
Order

Detection settings apply the next time you press Start.

Microphone

The model downloads once from Hugging Face and your browser caches it, so later visits start in seconds. Audio never leaves your device.

English only. Moonshine is trained on English speech; other languages come out as garbled English.

Transcript

Press Start and speak. Each sentence appears here after a short pause.

What this tool does

Live Transcription turns what you say into text while you speak. Press Start, talk into your microphone, and each sentence appears in the transcript a moment after you pause. It works like dictation or voice typing, but the speech recognition runs entirely in your browser: no account, no server, and no audio leaves your computer.

The recognition model is Moonshine, an open speech-to-text model from Useful Sensors designed for live, on-device transcription. It is fast on short phrases, which makes it a good fit for transcribing one sentence at a time. The tool runs it with Transformers.js and ONNX Runtime Web, on WebGPU when your browser has it and on WebAssembly otherwise.

This tool is English only.

How to use it

  1. Choose a model. Base is more accurate; Tiny is smaller and faster on slow computers.
  2. Click Download model and start. The model downloads once (about 120 MB for Base, 50 MB for Tiny), then your browser asks for microphone access. Allow it.
  3. Speak normally. The level meter shows the input volume and turns green while speech is detected. When you pause, the phrase is sent to the model and the text appears as a new line.
  4. Press Stop to release the microphone. Press Start again to continue; timestamps keep counting from where you stopped.
  5. Use Copy or Download .txt to keep the transcript, or Clear to start over. Set Order to Newest first to keep the latest line at the top; Copy and Download follow the same order.

Example

With timestamps on, a short dictated note looks like this:

[00:00] Let's start with the quarterly numbers.
[00:04] Revenue is up twelve percent compared to last year.
[00:09] Next, I want to talk about hiring.

Turn timestamps off to get plain lines, ready to paste into an email, a document or a note-taking app.

Tuning detection

A simple voice activity detector listens to the loudness of the microphone and decides where sentences begin and end. Three settings control it:

  • Speech threshold sets how loud a sound must be to count as speech. Choose a noisier setting if background noise keeps the meter green, or a more sensitive one if quiet speech is missed.
  • Pause that ends a line is how long you must be quiet before the phrase is transcribed. Shorter pauses give faster results; longer ones keep sentences together.
  • Longest line cuts very long stretches of speech so the transcript keeps moving even if you never pause.

Changes to these settings apply the next time you press Start. The model choice and the timestamp switch are saved in the page address, so you can bookmark or share your preferred setup. The transcript itself is never put in the link.

Model, license and limits

The tool uses onnx-community/moonshine-base-ONNX or onnx-community/moonshine-tiny-ONNX, ONNX conversions of the English Moonshine models, which are released under the MIT license. Transformers.js is Apache-2.0 licensed. Accuracy is best for clear speech from one speaker close to the microphone; names, rare words and heavy accents may be misheard, and punctuation is the model’s guess. For recorded audio files, use the Speech to Text tool instead.

FAQ

› Is my voice uploaded anywhere?

No. The microphone audio is captured, split into sentences and transcribed inside your browser tab. Only the model files are downloaded, from Hugging Face, the first time you press the download button. The transcript stays on the page and is not included in share links.

› Which languages does it understand?

English only. The Moonshine models used here are trained on English speech. Moonshine also has models for other languages, but they are released under a non-commercial license, so this tool does not use them. Speech in another language comes out as garbled English.

› How big is the download and do I need it every time?

The Base model is about 120 MB and the Tiny model about 50 MB. Your browser caches the files after the first download, so later visits load in a few seconds and work without downloading again.

› Which browsers work?

A recent Chrome, Edge, Firefox or Safari with microphone access. Browsers with WebGPU run the model on the graphics card; others use WebAssembly on the CPU, which is slower but still keeps up with normal speech on most laptops. The page must be allowed to use the microphone.

› Why does a line appear only after I stop talking?

The tool waits for a short pause before it transcribes what you said, because Moonshine works on whole phrases. You can make the pause shorter in the options, or lower the longest line length so long monologues are cut more often.