How to Transcribe Audio to Text — Without Uploading It

Transcribe a recording in four steps

Drop the file onto the transcriber: an iPhone voice memo (.m4a), an MP3, a WAV from a recorder, or a video — its soundtrack is used. Choose the language the recording is spoken in; the page guesses from your browser’s language, but a recording in another language needs it set, because the model is told what to listen for rather than detecting it. Leave the model on Standard for a first try, and click Transcribe.

Text appears every thirty seconds of audio, each line with the time it starts, so you can read the beginning while the rest is still being worked on. When it finishes, copy it or save it.

The first time: a one-off download

To run on your device, the speech model has to be on your device. The first transcription downloads it; your browser then keeps it, and later transcriptions start at once. The sizes, measured on 2026-10-08:

• Standard (Whisper base): about 80 MB when it runs on the processor, about 200 MB on the graphics card.
• Accurate (Whisper small): about 250 MB on the processor, about 590 MB on the graphics card.
• Plus about 6 MB, once, for the software that runs the model (ONNX Runtime and transformers.js, from the jsDelivr CDN).

How long it takes

Your device does the work, so its speed decides. Browsers that offer WebGPU — Chrome and Edge on most Windows and Mac computers — run the model on the graphics card. On a 2016-era desktop graphics card, thirty seconds of speech took about 17 to 25 seconds with Standard and about 26 with Accurate, including the start-up; a minute-long memo is done in about a minute.

Without WebGPU the model runs on the processor, in WebAssembly, on a single core: about two minutes per minute of audio with Standard on a fast desktop, and several times longer with Accurate. That is fine for a voice memo and slow for a two-hour meeting. The page says which of the two your browser will use before you start, and you can stop at any point and keep what is done.

Standard or Accurate?

Start with Standard. On a real 72-second English voice memo it got nearly every word, with sensible sentence breaks. Accurate, a model about three times larger, punctuated better and fixed the few words Standard misheard, and it is the better choice for Korean, Japanese, Chinese, strong accents and noisy rooms. An even smaller model was tried and left out on purpose: on the same memo it fell into a loop repeating two words for most of the transcript.

Getting a transcript you can rely on

Speech recognition is very good and not perfect. A few habits make the difference:

• Set the right language. A Korean recording transcribed as English comes out as an English translation or nonsense.
• Cut long silences and music first with the audio editor; the model can invent words in silence.
• Read names, numbers and technical terms against the recording — they are where mistakes cluster. The timestamps tell you where to listen.
• Expect one continuous transcript: it does not label who is speaking. Paragraph breaks follow long pauses, which often, but not always, fall between speakers.

Text, or subtitles

Four ways to save the result:

• .txt — the transcript as plain text, with or without timestamps (the Timestamps switch decides). Opens in anything, pastes into Word or Google Docs.
• .docx — the same text as a Word document, one paragraph per pause (or per line with timestamps). Opens in Word, Google Docs, Pages and LibreOffice.
• .srt — the most widely accepted subtitle format: YouTube, VLC, Premiere, DaVinci Resolve and most players and editors.
• .vtt — WebVTT, the subtitle format web video players use.

Why not just upload it?

Meetings, interviews, lectures, medical appointments and voice notes are among the most private files people have, and transcription services keep what they are given for some time — to process it, to show it back to you, sometimes to improve their models under their terms. Running the model on your device removes that question: there is no copy anywhere else. The trade is the download and your device’s speed, which is why this page tells you both before you start.

If it does not work

If the file will not open, it is in a format browsers do not decode — WMA, AIFF or AMR; convert it to MP3 with a desktop program first. If the model download stops, check the connection and try again; files that finished downloading are kept. If transcription stops with an out-of-memory error, close other tabs, or switch to Standard: Accurate on the graphics card needs a lot of video memory. On a phone, prefer Standard and short recordings.

Step by step

1. Drop the recording or video onto the transcriber.
2. Choose the spoken language, and Standard or Accurate.
3. Click Transcribe; the first run downloads the speech model once.
4. Copy the text, or save it as TXT, Word (.docx), SRT or VTT.

Frequently asked questions