Speech to text.
Audio stays here.
Record live, drop a file, or pick a take you already saved in the voice recorder. A small Whisper model does the work right in this tab. The model downloads once; your audio never leaves the browser.
Private transcription
One download, then it stays private
~40 MB Whisper model loads into this browser. Your audio never leaves the tab — only model weights hit the network, and only when you ask.
Loading the speech model you already approved on this browser…
Choose how you want to add audio.
Not hearing anything? Check your mic first. Want to trim first? Open the voice recorder.
Checking local library…
Editable text
Audio to text, without handing over the audio
Most online transcription pages are upload forms with a progress bar. This one works the other way around: decode locally, transcribe locally, copy when you choose. There is no server waiting for your clip.
Model download ≠ audio upload
The CDN only delivers Whisper weights after you opt in. Your mic stream and files stay in the tab.
Three private sources
Live record, dropped media, or takes already saved by the voice recorder in this browser.
You export on purpose
Copy or download TXT/SRT when you want the text. Nothing is shared silently.
Load once, then it just works
First use downloads a small English speech model. After that it loads from the browser cache — pick a source and go. No account at any point.
- 1
Load the model
Allow the ~40 MB download once. It usually caches in this browser for next time.
- 2
Choose audio
Record, drop a file under ten minutes, or select a local library take.
- 3
Edit and export
Fix names in the transcript box, then copy or download plain text / SRT.
Treat browser output as a strong first pass
Clean audio wins
One speaker, steady volume, and low room noise beat a bigger model almost every time. If the take is rough, trim it or check the mic first. Combining a few recordings into one? Prep or stitch recordings in the Audio Studio first.
Tiny is fast on purpose
Whisper tiny English keeps the download small and inference snappy in a tab. Hard accents, jargon, and overlapping talk still need a human pass — the editable box is there for that.
When you outgrow the tab
Long files, system-wide dictation, and Clabrate’s on-device Parakeet stack live in the Mac app. Same rule — your voice stays on your device — with more room to run.
Why not browser speech APIs?
Built-in recognition often leaves the device and cannot transcribe files you drop. One local engine keeps record, file, and library on the same private path.
Here, transcription ends in copy and paste. On a Mac it can just be typing.
A web tool always ends the same way: select, copy, switch windows, paste. The Clabrate Mac app removes that step — hold a key, talk, let go, and the text lands where your cursor already is, transcribed on your Mac by default.
This page exists so anyone can try private speech-to-text without installing anything. Plenty of people will use it and never need more — good. If you catch yourself dictating all day, that is the job the app was built for.
Transcription FAQ
Is this free audio-to-text converter private?
Yes for your audio. Recording, file decode, and transcription all run in your browser. Your voice is not uploaded to Clabrate. The only network download is the speech model (~40 MB) from a public CDN, and only after you choose Load private speech model.
What model does the browser use?
OpenAI Whisper tiny English (onnx-community/whisper-tiny.en) via Hugging Face Transformers.js. It runs with WebGPU when available, otherwise WebAssembly. The Clabrate Mac app uses a different on-device stack (Parakeet) for system-wide dictation.
Does loading the model upload my audio?
No. The model download only fetches model weight files. Your microphone stream, dropped files, and local library takes stay in this tab.
Can I transcribe a recording I made in the voice recorder?
Yes. Open the From library tab. Finished takes saved by the Clabrate voice recorder live in this browser’s local storage and can be transcribed without re-uploading.
What file types work?
Anything your browser can decode: typically MP3, WAV, M4A/AAC, OGG, FLAC, WebM, and many video containers for their audio track. If a file fails, convert to WAV or MP3 and try again.
Is there a length limit?
Yes — about 10 minutes per clip, so the tab stays responsive and your battery stays sane. Longer recordings are a better fit for the Mac app.
How accurate is browser Whisper tiny?
Strong on clean, single-speaker English. Names, heavy accents, crosstalk, and noisy rooms still need a human pass. Treat the output as a first draft you can edit in the transcript box.
Why not use the browser’s built-in speech recognition?
Built-in SpeechRecognition often sends audio to the browser or OS cloud service, only works for live mic (not files), and is weaker on names and jargon. This tool keeps one private local engine for record, file, and library paths.