How to transcribe audio without uploading it

Most transcription sites ask you to upload the recording first. That is a poor fit for interviews, meetings, lectures or voice notes you would rather keep to yourself. A speech model can run inside your browser instead. TypeMaster downloads the model once, then turns your audio into text on your own device. There is no account to create.

Open Speech to text

Transcribe audio in four steps

Choose your audio. Pick or drop an MP3, WAV, M4A or OGG file, or a video such as MP4 or MOV. You can also press Record with microphone and allow access when your browser asks. If you refuse, nothing is recorded and you can still choose a file.
Pick a model and language. The size of the download is shown before anything is fetched. If you know the spoken language, choose it rather than relying on detection.
Press Transcribe. Progress reads like “Transcribing 2:30 of 10:00”. Cancel stops the work immediately. The first run also downloads the model, with a progress bar.
Review and save. Fix mistakes in place, switch timestamps on or off, then copy the text or download TXT. For captions, load the same file in the Subtitle maker.

Tiny or base: which model?

Both options are sizes of OpenAI’s Whisper, shrunk with quantization so they fit in a browser. Accurate (base) is the default. It makes fewer mistakes with accents, background noise and uncommon words, but it downloads about 96 MB on most devices and works more slowly. Faster (tiny) downloads about 60 MB and transcribes noticeably quicker. It suits clear speech from one person, rough drafts and older phones.

A practical approach: start with tiny on a long file, and switch to base if the errors get in your way. A model you have downloaded stays in your browser, so trying the other one later costs only that model’s download. The page always shows the exact size for your device.

Tips for better accuracy

Record close to the speaker and cut background music or street noise where you can.
Choose the spoken language. Detection can guess wrong on short clips or strong accents.
Keep one voice at a time. Overlapping speakers are the hardest case for any model.
Check names, numbers and technical terms yourself; these are the usual mistakes.
Delete invented words that appear over music or long silences.

What stays on your device

Your browser decodes the recording and passes it to a background worker that runs the model. Your audio and your transcript are never sent anywhere. What does travel over the network is the model itself: its files are fetched once from Hugging Face, a public model host, together with this site’s own files. Your browser keeps them, so later runs work offline.

You do not have to take that on trust. Open your browser’s Network panel, transcribe a harmless sample, and look at what leaves your device. You should see requests that download files, and none carrying your recording. How to check if a PDF tool uploads your file walks through the method; the steps are the same for audio.

Limits to know

Files can be up to 60 minutes and 500 MB. Your browser must be able to read the format, and very long recordings take time and memory on a phone, so close other tabs first. A transcript is a draft. Review it before you rely on it for anything that matters.

Video works when your browser can read its audio track, as it can for most MP4, MOV and WebM files. If a file will not open, export the audio as MP3 or WAV from any audio or video app, then try again. Recordings made in the page are saved nowhere: they are transcribed from memory, and closing the tab discards them.